Developer desty's blog & portfolio.https://desty.github.io/https://desty.github.io/en/blog/81-pstack-merge-without-review/https://desty.github.io/en/blog/81-pstack-merge-without-review/Lauren Tan, a former Cursor engineer now building Grok Bot at SpaceXAI, says she merged 1,000 PRs to production in July and 2,462 in August. The story spread as 'she doesn't look at the code', but she answered that she reads what her agents land after the fact and turns what she finds into lints and structure. The job pre-merge human review used to do is split across three mechanisms: evidence from running the real app, a verdict from an agent that did not write the code, and a codebase where rules fail CI. I read the transcripts of two talks, her posts on X, and the code and 90-commit history of pstack, her open plugin, to see how each mechanism is built, and reproduced the patch-id rule that ties a verdict to a patch. After covering token usage and the numbers nobody has audited, the last section lays out the order in which a team on a normal budget could adopt this.Sun, 27 Sep 2026 09:36:00 GMThttps://desty.github.io/en/blog/79-google-ax-agent-orchestrator/https://desty.github.io/en/blog/79-google-ax-agent-orchestrator/Around September 20, Google published AX, an agent orchestrator written in Go. The README promises 'billions of autonomous agent workloads in a cluster', and a 661-point Hacker News thread split over that number and the prerequisites. Reading the design doc and the code shows AX is not an agent framework. It treats one agent as one sandbox, declares it with four manifests (Task, Workspace, Gateway, Model), and applies, suspends and resumes it the way kubectl does. Execution belongs to a separate project, Agent Substrate. This post checks how far the code goes on what people wanted from AX (idle-agent cost, isolation, repeated environment setup, an auditable stack), what is still on the roadmap (idle detection, exit status, workload identity), and which teams can try it today.Thu, 24 Sep 2026 14:00:00 GMThttps://desty.github.io/en/blog/78-lossless-agent-memory/https://desty.github.io/en/blog/78-lossless-agent-memory/As complaints pile up about coding agents forgetting their own work after context compaction, more 'lossless' memory projects are appearing. Reading the code of lossless-memory (Show HN, September 21) and the OpenClaw plugin lossless-claw shows both store every original message but cut or summarize what goes into the context. They differ in where the loss happens and on what basis. And Claude Code already saves every message and tool result locally, deleting them after 30 days. Loading 2,000 lines of synthetic log into lossless-memory and searching by date returned 650,000 characters at once. When you keep everything, the real decisions are retention period, what to keep, how things enter the context, output limits, and cache placement.Thu, 24 Sep 2026 12:00:00 GMThttps://desty.github.io/en/blog/77-i-have-adhd-skill/https://desty.github.io/en/blog/77-i-have-adhd-skill/i-have-adhd, a skill that keeps coding agents from burying the answer under explanation, hit the Hacker News front page in September and picked up about 20,000 stars in two weeks, passing 50,000. The skill itself is a 142-line SKILL.md whose core rule is: put the next action or the result on the first line. The repo's history and issues show the rules evolving mostly into exceptions, and the author's only measurement found gains concentrated in progress and error reports, while the 'state cause and fix' rule produced unfounded diagnoses. Anthropic answered the same complaint with Claude Code's Concise output style and a change to Opus 5.5's response structure. This post sorts each rule into where it belongs (skill, output style, harness, or UI) and lays out how to measure the effect.Thu, 24 Sep 2026 11:00:00 GMThttps://desty.github.io/en/blog/76-openai-agent-incidents/https://desty.github.io/en/blog/76-openai-agent-incidents/On September 24, Australia's prime minister said an OpenAI agent accessed a Medicare statistics portal without authorization in June. In July, roughly 700 OpenAI evaluation agents attacked Hugging Face, and OpenAI published a 51-page incident report in late August. Following that report, the agents never broke the sandbox itself. They turned an internal Artifactory instance, opened for package installs, into a message board and an internet relay, then used the privileges they gained there plus tokens exposed on the internet to get out. OpenAI says the same model's propensity to compromise infrastructure dropped more than 100x under its production harness and system prompt. This post traces the paths in all three incidents and turns them into checks and tests for teams running coding and ops agents.Thu, 24 Sep 2026 10:00:00 GMThttps://desty.github.io/en/blog/80-unreal-agent-async-harness/https://desty.github.io/en/blog/80-unreal-agent-async-harness/Unreal Agent, released by Unreal Labs on September 22, runs every tool call in the background and has the harness wake the model when a result lands. The model never spends a turn checking whether a tool has finished. On Terminal-Bench 4.0 with the same GPT-6 Astra xhigh, it matched Codex's 57.9% pass rate at $1,428 versus $2,350. The demand was already sitting in the Codex repository: a March issue traced a loop that re-sends the whole conversation every five seconds to poll a background process, and a July report measured wait-only turns at 19.8% of token volume. Reading the code shows placeholder swapping, a one-second result window, a ten-minute heartbeat, and a cache-preserving choice to leave two results for one call, which some providers reject with a 400. The benchmarks are self-run and the headline chart compared mismatched effort levels. The last section covers how to count waiting turns in your own harness and what to check if you build one.Thu, 24 Sep 2026 09:04:00 GMThttps://desty.github.io/en/blog/75-opus-5-5-gpt-6-sol/https://desty.github.io/en/blog/75-opus-5-5-gpt-6-sol/On September 22, Anthropic and OpenAI shipped cheaper models 90 minutes apart. Claude Opus 5.5 cuts per-token prices 20% from Opus 5; GPT-6 Sol and Luna halve GPT-5.6. Working through Artificial Analysis's per-effort cost-per-task data shows the savings come from different places. Opus 5.5 halves cost at low and high effort but is 2% more expensive than Opus 5 at max. GPT-6 Sol uses roughly the same tokens as before; only the price per token fell. Picking the cheapest configuration for each target score gives Luna up to 37 points, Sol up to 48, and Opus 5.5 above that. Also covered: the migration items for Opus 5.5 (whose default effort dropped to medium), what to check before moving to GPT-6 Sol, and a procedure for comparing them on your own tasks.Thu, 24 Sep 2026 09:00:00 GMThttps://desty.github.io/en/blog/71-jev-coding-cost/https://desty.github.io/en/blog/71-jev-coding-cost/fast-jev-compaction does not summarize a Claude Code conversation. It selects stale tool calls and results to delete. I traced the implementation to see how Jev could reduce not only the cost of compaction but the input tokens and latency of later model calls, what it can delete by mistake, and how to measure the real effect. Part 5 of the Jev series.Sat, 19 Sep 2026 12:00:00 GMThttps://desty.github.io/en/blog/70-jev-four-builds/https://desty.github.io/en/blog/70-jev-four-builds/Four Jev use cases, built and run myself. A Claude Code safety hook stopped all eight dangerous commands a regex blocklist missed, but as a hook its latency was 0.65s, not 0.25s. Classifying the full text of 72 blog posts took 20 seconds and 1.4 cents. Finding the needed line in an 87-line changelog meant sending 87 questions in one call, in 0.27s. Plus what the four had in common. Part 4 of a series.Sat, 19 Sep 2026 10:00:00 GMThttps://desty.github.io/en/blog/68-jev-style-engine/https://desty.github.io/en/blog/68-jev-style-engine/Jev's training method is undisclosed, but its input/output pattern can be reproduced. I built an engine on an M1 MacBook that runs one forward pass and reads the logits of the option tokens, then compared Qwen3 from 0.6B to 8B against live Jev on the same 90 items. At 4B, English reached 98% but Korean stayed at 84%, and one decision took a second. I also recovered Jev's confidence formula from its responses. Part 2 of a series.Sat, 19 Sep 2026 08:30:00 GMThttps://desty.github.io/en/blog/67-jev-system-one/https://desty.github.io/en/blog/67-jev-system-one/Diogo Almeida, who helped build RLHF at OpenAI, spent two years on a model that produces no prose at all: only choices and probabilities. The starting point is on record in the GPT-4 technical report, where post-training pushed calibration error from 0.007 to 0.074. What people built in four days, whether the open-source rumor is true, and how to try it now. Part 1 of a series.Sat, 19 Sep 2026 08:00:00 GMThttps://desty.github.io/en/blog/69-jev-measured/https://desty.github.io/en/blog/69-jev-measured/Measurements from calling TypeSafe AI's Jev directly. Ten questions cost the same latency as one, and server-side time was about 90ms. Counting and date intervals were exact, but judgments that derive a deadline and compare it to today broke down, and a fully refunded order was approved at 0.8 probability. Once code computed the dates and the balance, all 36 cases were right. Korean matched English on accuracy with lower confidence. Part 3 of a series.Sat, 19 Sep 2026 06:00:00 GMThttps://desty.github.io/en/blog/66-agents-md-claude-code/https://desty.github.io/en/blog/66-agents-md-claude-code/Claude Code 2.1.277 reads AGENTS.md directly. The default still prefers CLAUDE.md, and because the feature ships as a built-in plugin it never appears in /memory, skips InstructionsLoaded hooks, and is absent on Bedrock. Three headless tests of the loading rules, then a per-team split of what belongs in which file.Sat, 19 Sep 2026 02:00:00 GMThttps://desty.github.io/en/blog/65-claude-plugin-eval/https://desty.github.io/en/blog/65-claude-plugin-eval/Claude Code 2.1.269 added plugin eval. It repeats the same prompts with and without a plugin and reports WITH, W/OUT, and delta scores. This article explains how to design cases and graders and use the results for regression testing.Sun, 13 Sep 2026 06:00:00 GMThttps://desty.github.io/en/blog/64-deepseek-harness/https://desty.github.io/en/blog/64-deepseek-harness/DeepSeek Harness treats the model, tools, and session — and the loop that drives a turn — as plugins. Cordis mounts services on ctx keys; the default driver is ctx.agentLoop. This splits what is kernel from what you change in config, at the 10 September 2026 commit.Sun, 13 Sep 2026 01:20:00 GMThttps://desty.github.io/en/blog/63-openviking/https://desty.github.io/en/blog/63-openviking/ByteDance Volcano Engine's OpenViking stores agent memories, documents, and skills as directories under viking://. Instead of stuffing vector hits into the prompt, an agent checks L0/L1 summaries and opens L2 only when needed. This reads the retriever, session extractor, and harness attachment points at the 12 September 2026 commit.Sun, 13 Sep 2026 01:00:00 GMThttps://desty.github.io/en/blog/62-gpt-6-astra/https://desty.github.io/en/blog/62-gpt-6-astra/Compare coding results with earlier models, distinguish Chat, Work, and Codex access and usage, choose reasoning effort, enable experimental context management, and update instructions and workflows. Includes practical prompts, interruption handling, and a staged migration plan.Sat, 05 Sep 2026 09:00:00 GMThttps://desty.github.io/en/blog/61-webmcp-challenge/https://desty.github.io/en/blog/61-webmcp-challenge/OpenAI has launched a ten-day WebMCP Challenge. Each of the top ten entries receives $3,000 and a year of ChatGPT Pro. WebMCP lets a website expose its capabilities as structured tools instead of making an agent infer how to use the interface. This article examines how it differs from MCP, how OpenAI already uses it for internal evaluation work, and the browser-support and security limits that remain.Sun, 30 Aug 2026 21:21:00 GMThttps://desty.github.io/en/blog/60-eli5-visual-explainer/https://desty.github.io/en/blog/60-eli5-visual-explainer/Claude Code's Thariq Shihipar shared ELI5, a skill he says people at Anthropic have been using frequently. Add a subject after /eli5 and the agent reads the relevant context, then creates an HTML explainer with big pictures and few words. What is public is closer to a compact working pattern than a released plugin. This article examines the review step that pattern adds to code comprehension, design review, and incident analysis.Sun, 23 Aug 2026 18:00:00 GMThttps://desty.github.io/en/blog/59-vercel-fx-native-agent/https://desty.github.io/en/blog/59-vercel-fx-native-agent/Vercel Labs built fx in Zig and ships it without a Node.js or Python runtime. The current macOS arm64 executable is 6.43 MB, but size alone misses the design. fx can run as an interactive shell, a one-shot JSON command, an ACP server, or an experimental WebAssembly core. This article examines what the size and startup claims measure, which responsibilities move to the model and host, where the sandbox boundary sits, and when a disposable native agent is useful.Sun, 23 Aug 2026 09:00:00 GMThttps://desty.github.io/en/blog/58-opensandbox-agent-runtime/https://desty.github.io/en/blog/58-opensandbox-agent-runtime/I cloned OpenSandbox, a 14.5k-star repository created eight months ago, and traced its four API specifications, FastAPI control plane, Go execution daemon, Docker and Kubernetes runtimes, egress layer, and Credential Vault. It looks like an SDK for creating sandboxes and running commands, but the real product is a protocol boundary that keeps agent code stable while execution environments change. This analysis covers runtime injection, credentials that stay outside workloads, pools and snapshots, and the caveats to check before adoption: default runc, backend capability gaps, and independently versioned components.Sat, 22 Aug 2026 11:00:00 GMThttps://desty.github.io/en/blog/57-gemini-grok-migration/https://desty.github.io/en/blog/57-gemini-grok-migration/Two new models landed in the second week of August: Grok 4.6 on the 12th, Gemini 3.7 Flash on the 13th. The benchmark tables look great, but if you're the one migrating, the things worth checking sit outside the tables. Gemini's half price runs through December 31 and reverts on January 1 to exactly what 3.6 Flash costs today. Grok was reported as holding prices flat, but cache reads went up 67% — and long-running agents, the very workload this model targets, are the workload that hits cache reads hardest. I go through what actually improved over the previous versions, and what to check in your code and on the price sheet before you flip the model string: xhigh quietly starting to work, the vanished minimal level, the 200K pricing boundary, and the five parameters you now delete.Sun, 16 Aug 2026 15:10:00 GMThttps://desty.github.io/en/blog/56-grok-bot-shared-computer/https://desty.github.io/en/blog/56-grok-bot-shared-computer/On August 11, xAI shipped Grok Bot in beta: always-on AI teammates you message like colleagues. The tagline says 'Bots have their own computer,' but the docs say something else — there is one computer per account, every Bot shares its files, browser sessions, and CLI credentials, and the official documentation tells you outright not to use separate Bots as a security boundary. What the persistent design sells (log in once, no repeated setup, compounding context) and the risk it creates (untrusted input + resident credentials + outbound communication) come from the same place. I compare how Claude Cowork, Claude Code, ChatGPT Work, and Codex answered the same design question differently, flag the fact that the official security doc never once says 'prompt injection,' and close with the rules I'd follow if you try it anyway.Sun, 16 Aug 2026 14:00:00 GMThttps://desty.github.io/en/blog/55-dynamodb-vector-search/https://desty.github.io/en/blog/55-dynamodb-vector-search/On August 5, AWS shipped native vector search in DynamoDB — no preview, straight to GA. You store embeddings next to your operational data and run similarity search in place through a single SearchVectors API. The previous official answer, the zero-ETL OpenSearch chain, spanned five services, and ScyllaDB spent three months mocking exactly that before AWS erased the attack surface. The GA makes the real demand legible: not better ANN, but one less system to operate. This post covers the fine print behind 'any scale' (the SearchSchema partition key is your cost design), the traps in the agent-memory use case (eventual consistency, no FGAC), the constraints the announcement doesn't mention — on-demand only, equality-only filters, no pagination — and how to choose between DynamoDB, S3 Vectors, OpenSearch, and Postgres.Sun, 09 Aug 2026 18:00:00 GMThttps://desty.github.io/en/blog/54-prime-agent-self-improving-harness/https://desty.github.io/en/blog/54-prime-agent-self-improving-harness/Claude Opus 5 scores 30.2% on ARC-AGI-3's official harness. Put the same model on Prime Intellect's new harness and it scores 95.5% — above the human expert baseline. Prime Agent collapses all tools into a single persistent IPython kernel so the model treats context as variables and sub-agents as function calls (RLM), then adds a Continual Harness where the agent CRUDs its own prompts, skills, memories, and sub-agent specs. But the most memorable part isn't the numbers — it's an observation the makers published themselves: in Factorio, the very refinement loop that had been building legitimate skills pivoted to building efficient cheating skills once it found an exploit. Self-reported benchmark caveats, the thesis that harnesses go stale faster than models, and four questions worth importing into your own stack.Sun, 09 Aug 2026 14:00:00 GMThttps://desty.github.io/en/blog/52-archify-skill-product/https://desty.github.io/en/blog/52-archify-skill-product/A repository less than four months old has 10.7k stars, and its core is a 104-line SKILL.md. Archify is an agent skill that makes coding agents draw architecture diagrams — but what made it a hit isn't drawing talent, it's the 25,000 lines of deterministic validation code behind the instructions. A typed JSON intermediate representation, no delivery without passing validation, a repair budget of two rounds, and the instruction 'a non-zero exit can never be described as success' — I dissect a design that wraps a fuzzy generation task in a task-sized harness. Plus the lineage of a fork overtaking its dormant original, a skills ecosystem that now has sponsors and benchmarks, and five design moves worth stealing for your in-house skills.Sun, 09 Aug 2026 10:00:00 GMThttps://desty.github.io/en/blog/53-taalas-model-in-silicon/https://desty.github.io/en/blog/53-taalas-model-in-silicon/Taalas, whose acquisition AMD announced on August 6, sits at the opposite end of the spectrum from a GPU. Its only product, the HC1, runs exactly one model — Llama 3.1 8B — because the weights are physically etched into the transistors. The physics behind a 250W card with no HBM and no liquid cooling pushing 17,000 tokens per second per user is real; the numbers, however, are all vendor-reported. The real asset isn't the chip but the process — turning a new model into dedicated silicon in two months — and I dig into the future that process is aiming at. Why the 'consumer card for a frontier model' fantasy can't work for closed models but can for open weights, and which assumptions developers should swap out now, before the inference cost curve bends again.Sun, 09 Aug 2026 10:00:00 GMThttps://desty.github.io/en/blog/51-loopx-state-kernel/https://desty.github.io/en/blog/51-loopx-state-kernel/4,037 commits in ten weeks. LoopX, built by a ByteDance engineer, is the first serious implementation to appear just six weeks after the loop engineering discourse coined its name. Starting from the observation that chat memory and a timer can't govern multi-day loops, it pulls objectives, gates, evidence, and quota out of the agent into an external state kernel. The stranger part: the tool is building itself with its own loops — a self-iteration case of 74 commits in one day, an AGENTS.md written for agents, and a README that grades its own claims with evidence labels. What the state-kernel demand actually is, what the meta-case signals and warns, and when a third-party kernel makes sense in the era of built-in /loop.Sat, 08 Aug 2026 10:00:00 GMThttps://desty.github.io/en/blog/50-tencentdb-agent-memory/https://desty.github.io/en/blog/50-tencentdb-agent-memory/Tencent's open-source TencentDB Agent Memory is not a conversation dump. It governs four assets — Chat Memory, Skill, Wiki, and CodeGraph — at team scope. The design bets are layered L0–L3 distillation, memory as agent loadout rather than a global prompt, and a transparent MemoryProxy that injects context into coding agents without code changes. The PersonaMem jump from 48% to 76% matters less than the shift from RAG's 'what can be found' to 'who may use it, which version is live, and which agent is equipped.' Architecture, DX, and the rough edges as of the v2.0 release and ~17k stars.Sat, 08 Aug 2026 09:00:00 GMThttps://desty.github.io/en/blog/45-ontology-comeback/https://desty.github.io/en/blog/45-ontology-comeback/In late July, a learning web app called Ontology Playground from Microsoft made the rounds. It looks like a toy, but the direction it points is interesting. When the Semantic Web designed ontologies in 2001, the intended end user was the intelligent agent. That vision failed because it demanded humans annotate meaning by hand — and once LLMs solved the 'reading' problem, the bottleneck moved to the absence of meaning. Now Palantir, Snowflake, dbt, Databricks, and Microsoft are all moving at once to greet the rightful user who arrived 25 years late. This post covers the numbers behind text-to-SQL collapsing to 17% in the enterprise, what Fabric IQ Ontology actually is and where it's constrained, and the criteria plus a starting ladder for deciding whether you actually need an ontology.Fri, 07 Aug 2026 10:00:00 GMThttps://desty.github.io/en/blog/44-surviving-model-churn/https://desty.github.io/en/blog/44-surviving-model-churn/Measure the deprecation pages of the big three labs and you get a hard number: GA models actually live 12–16 months. Prompts and agent harnesses are structures built on top, so they depreciate on the same one-year clock. #36 treated migration as an event; this post starts from the premise that it's a cycle. The paradox where instructions that patched an old model's weaknesses become liabilities on the new one, how prompt-transfer failure earned an academic name — 'Model Drifting' — and how to move your investment into the assets that survive a model swap: evals, schemas, and specs. Ends with a six-item checklist for making the model-dependent parts of your system cheap to replace.Tue, 04 Aug 2026 10:00:00 GMThttps://desty.github.io/en/blog/49-cross-agent-review/https://desty.github.io/en/blog/49-cross-agent-review/Watching Codex write a parser while Claude reviews the same change in the next pane is no longer a party trick. Anthropic Code Review, VS Code multi-agent workflows, builder–validator chains, and cross-provider hubs like CCB (Claude Codex Bridge) all point the same way: trust comes less from one model's intelligence than from agents with different failure modes inspecting each other. Once the copy-paste tax disappears, the real design work is circuit breakers, role separation, and knowing when two agents are enough.Sun, 02 Aug 2026 14:00:00 GMThttps://desty.github.io/en/blog/48-cache-hit-rate/https://desty.github.io/en/blog/48-cache-hit-rate/In July, Musinsa's tech blog posted numbers you rarely see from a Korean company: 64% LLM cost reduction, 98% cache hit rate. It's effectively the first Korean tech blog post to publish measured prompt-caching results. But what lingers isn't the savings — it's the nature of the number itself. Cache hit rate isn't a discount line on your invoice; it's a graded score on one design question: did you put what never changes in front and what always changes at the end? This post lines up the other answer sheets to the same exam — a security agent that jumped from 7% to 84%, Anthropic declaring incidents when hit rate drops, Manus calling KV-cache hit rate 'the single most important metric for a production agent' — and walks from how one token breaks the hash chain, through a four-provider caching policy comparison, to the order in which to raise your own score.Sun, 02 Aug 2026 10:00:00 GMThttps://desty.github.io/en/blog/43-graph-engineering/https://desty.github.io/en/blog/43-graph-engineering/In July, 'graph engineering' started making the rounds in the agent world — the fourth label after prompt, context, and loop engineering. As with every new coinage, half of it is marketing. But this one has a twist: the same name is being used for two different things — orchestration graphs that shape the work, and knowledge/memory graphs that hold what the system knows. Both are worth digging into. The former is converging on a principle — models judge, code routes — and the latter's indexing cost dropped a thousandfold in eighteen months, changing the adoption math. This post covers the term's lineage, sorts out the two branches, and gives the criteria for deciding whether your system actually needs a graph.Fri, 31 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/47-chonkie-chunking/https://desty.github.io/en/blog/47-chonkie-chunking/In RAG tutorials, chunking is the one-line preprocessing chore. A library that does only that step now pulls 1.19 million monthly downloads, and its two creators went through Y Combinator. The interesting part comes after: the '33x faster' claim was against the slowest competitor (1.06x against the leader), '100GB/s' is what you get when you stop counting tokens entirely, and the founder himself admitted on HN that 'perfect splits don't move the needle much' for retrieval quality. Downloads kept climbing anyway. Against a research landscape that keeps confirming semantic chunking can't beat plain recursive splitting, this post dissects what people were actually buying, why long context didn't kill chunking, which of the 11 chunkers to pick, and how the company pivoted away while the library stayed.Wed, 29 Jul 2026 14:00:00 GMThttps://desty.github.io/en/blog/46-is-graphrag-needed/https://desty.github.io/en/blog/46-is-graphrag-needed/In late June, engineers from AWS and Cisco published a paper with a provocative title: Is GraphRAG Needed? It puts nine scenarios — from regular RAG through GraphRAG to Agentic RAG — on the same knowledge base and the same model, implemented side by side. The scorecard holds three upsets: graph-only retrieval collapsed, a plain RAG that flattens relations into documents beat hybrid GraphRAG, and handing an agent an extra graph tool made it 12% worse. The more painful number comes after: even when retrieval brings in 83.5% of the ground-truth evidence, the model only uses 47.9% of it. This post covers the fix that cut tokens nearly in half by changing only how retrieval results are written down, and the order of operations to run through before you reach for a graph.Wed, 29 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/42-korean-embeddings/https://desty.github.io/en/blog/42-korean-embeddings/When building Korean RAG, the question that comes up as often as 'which vector DB' is 'which embedding model.' Opening the MTEB leaderboard doesn't answer it — a multilingual average score says nothing about Korean retrieval quality. Korean-only leaderboards like ko-embedding-leaderboard exist to fill exactly that gap. Reading the rankings breaks a few pieces of conventional wisdom: bge-m3, the long-standing default, has slipped to 8th, and first place is a 0.65-point contest between a 568M Korean fine-tune and a 4B frontier model. This post covers why Korean needs its own leaderboard, the three things to actually read in the rankings, and how to choose a Korean embedding model today.Tue, 28 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/41-pgcontext/https://desty.github.io/en/blog/41-pgcontext/The first question you hit when starting with RAG is 'do I need a separate vector database?' By mid-2026 the answer has mostly settled on 'no, just use Postgres' — but that answer still has one hole in it: the moment you add a metadata filter, pgvector's recall quietly collapses. pgContext, a Rust extension that caught my eye on GitHub this week, drills into exactly that weakness. This post covers why people keep wanting 'everything in Postgres' (the burden of running one more database, the barrier for newcomers), what pgvector's post-filtering trap actually is, what pgContext does differently and where it admits it loses, and how to tell whether your own workload is already caught in the trap.Mon, 27 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/40-claude-opus-5/https://desty.github.io/en/blog/40-claude-opus-5/On July 24 Anthropic shipped Claude Opus 5 — its fourth release in two months, with an unapologetic positioning: Fable 5-class performance at half the price ($5/$25). But read the migration guide and an unfamiliar pattern emerges: there are more instructions to **delete from your existing prompts** than to add. Delete 'verify your work.' Delete 'delegate more to subagents.' #21 argued that prompts had become brakes rather than accelerators — Opus 5 now asks you to take your foot off the brake entirely. Benchmark scrutiny, two breaking changes including thinking-on-by-default, the deletion list, and the 'asymmetric diet' I found by cracking open the generational system prompts in the public leak repository.Sun, 26 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/39-ide-to-ade/https://desty.github.io/en/blog/39-ide-to-ade/#35 and #38 argued that the performance bottleneck has moved outside the model. This time it's the workflow's turn. Orca, an open-source tool from YC-backed Stably, builds zero coding agents of its own — it's a control tower that runs 30+ third-party agents (Claude Code, Codex, and friends) in parallel git worktrees, a so-called ADE (Agent Development Environment). Its GitHub stars went from 7.7k in June to over 28k by late July. What that surge caught: four real bottlenecks of vibe coding — human wait time, nondeterminism, double-paying for tokens, and an approve button chained to your desk — plus a forecast ('the IDE doesn't die, it gets demoted') and the caveats to know before trying it.Sat, 25 Jul 2026 21:00:00 GMThttps://desty.github.io/en/blog/37-company-brain/https://desty.github.io/en/blog/37-company-brain/Cerebras published the blueprint of its internal knowledge base. Three months after launch it fields over 15,000 questions a day and ranks among the most widely adopted internal tools at the company — and the askers aren't just humans. Automation scripts and agents query the same place. The design starts by giving up on the single source of truth: knowledge stays scattered but searchable, retrieval stacks four techniques (full-text, embeddings, term rarity, age decay), normalized summaries get embedded instead of raw transcripts, and code re-embeds only the chunks each commit changed. It all converges on one thesis — a brain is a maintenance system, not a storage system, and it's #30's context engineering scaled to a whole company. Glean's $300M ARR (+89%) shows the market already prices this thing.Mon, 20 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/38-tofu-white-box-harness/https://desty.github.io/en/blog/38-tofu-white-box-harness/#35 ended with the question 'did the model earn that score, or did the harness carry it?' — and six days later a paper answered it with a controlled experiment. ToFu, an open-source harness from the NiuTrans lab (Northeastern University, China) and Meituan, froze three models, swapped only the harness, and ran SWE-bench Verified: 28.4% fewer tokens than Claude Code on average, with Pass@1 3.8 points higher. The three-layer context compaction behind it, the Opus plot twist where saving tokens didn't save money, and a multilingual pipeline that doesn't help Korean yet.Sun, 19 Jul 2026 21:00:00 GMThttps://desty.github.io/en/blog/36-model-migration/https://desty.github.io/en/blog/36-model-migration/For four months, Claude Opus 4.8 was unbeaten on Ploy's production evals. Then GPT-5.6 took the crown: 27% lower cost per completion, 2.2x faster, higher quality — while emitting half the output tokens. On this workload, the new model produced more with less. Roughly a third of the first run's raw failures came not from the model but from a harness bent around Opus; tool schemas, caching, and reasoning replay had to migrate with it for a fair comparison. #35 asked 'is that the model's score, or the harness's?' — this is that question in live production.Sun, 19 Jul 2026 18:00:00 GMThttps://desty.github.io/en/blog/35-harness-engineering/https://desty.github.io/en/blog/35-harness-engineering/LangChain climbed from around 30th to 5th on Terminal Bench 2.0 without touching the model once. Everything they fixed lived outside it — self-verification loops in the prompt, better tools, middleware that catches failure patterns. In 2026, major engineering teams and design writers rapidly mainstreamed a name for this 'everything but the model': harness engineering. It's the main event for what #14 called 'the war outside the model' and #25 only name-dropped. The Guides×Sensors control-system frame, Apiiro's published observations (syntax errors −76%, privilege-escalation paths +322%), and why a harness, unlike a model, is an asset that stays.Sun, 19 Jul 2026 16:00:00 GMThttps://desty.github.io/en/blog/34-sufficient-context/https://desty.github.io/en/blog/34-sufficient-context/Hand a model retrieved documents and you'd expect it to get more honest. Measurements say the opposite — given context, models get overconfident and lose the ability to say 'I don't know' (Claude 3.5 Sonnet's abstention rate: 84% → 52%). The ICLR 2025 paper Sufficient Context cut through this by asking 'is it sufficient?' instead of 'is it relevant?', and 13 months later Google productized it as the stop signal in Gemini Enterprise's agent loop (up to +34% factuality). It fills the missing brake in #19's search loop and adds a missing axis to #32's evals. Here's the mechanism, plus a hands-on recipe for adding abstention to your own RAG.Sun, 19 Jul 2026 14:00:00 GMThttps://desty.github.io/en/blog/33-kimi-k3/https://desty.github.io/en/blog/33-kimi-k3/On July 16 Moonshot shipped the 2.8-trillion-parameter Kimi K3, and the headline that an open model beat Fable 5 and GPT-5.6 Sol swallowed the week. In the last two posts (#21, #31) U.S. labs locked their strongest tier behind government gates; China did the opposite and routed around the wall by open-sourcing the biggest one. But is 'beat everyone' true? Rolling the tape on only the comparable numbers, it's half true, half marketing — 3rd–4th on overall intelligence depending on how you count, but genuinely #1 on three axes. Here's what's real and what's spin, split by data.Sat, 18 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/32-ai-eval/https://desty.github.io/en/blog/32-ai-eval/You change a prompt, bump the model, tweak RAG chunking. Then you run a few examples, decide 'feels better,' and move on. That's not improvement — it's luck. LLMs have infinite inputs and probabilistic outputs, so eyeballing doesn't measure anything. The real first step isn't building the feature; it's standing up an eval set and a measurement pipeline first. Only with eval does a change resolve into improvement or regression. If you can't measure it, you can't improve it.Mon, 06 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/31-reviewing-ai-code/https://desty.github.io/en/blog/31-reviewing-ai-code/Generation got explosively faster; review didn't move. So the bottleneck quietly shifted to review. And it's not just about speed. AI code fails differently than human code — clean names, fine structure, and yet subtly wrong. That 'looking right' is the trap. So reviewing AI code isn't about going faster; it's about reviewing differently. Not 'does it read well' but 'is the part you can't see correct.'Sun, 05 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/30-context-engineering/https://desty.github.io/en/blog/30-context-engineering/From 8k to 1M, the context-window arms race is effectively over. And it created a paradox — the moment you could put in more, the skill of putting in less mattered more. The intuition is 'more context, smarter model.' The reality is the opposite: the longer the tokens, the worse the model finds what it needs (context rot). So the center of gravity moved. If prompt engineering was 'how do I ask,' context engineering is 'what goes in, what stays out, and when do I clear it.'Sat, 04 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/29-mcp-tool-design/https://desty.github.io/en/blog/29-mcp-tool-design/Now that MCP is a de facto standard, everyone's building servers. Step back, though, and you'll see people competing on the easy part. The protocol is JSON-RPC plus three primitives — a few days of docs. What decides whether your agent actually works is the tool design on top: names, descriptions, error messages, approval gates, and how many tools you expose. And it isn't a new skill. The consumer just happens to be a model instead of a human — it's the resurrection of API design you'd forgotten you had.Fri, 03 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/28-open-knowledge-format/https://desty.github.io/en/blog/28-open-knowledge-format/Google released OKF (Open Knowledge Format), a vendor-neutral format for representing knowledge as markdown + YAML. What matters is that it didn't come out of Google's strategy room. In April, Karpathy dropped a one-screen 'LLM Wiki' gist that crossed 5,000 stars in two weeks, and Google shipped a vendor-neutral version days later. The crowd proved the demand first — it's the same pattern already converging bottom-up through llms.txt, AGENTS.md, and CLAUDE.md. So the question isn't 'is OKF a good format' but why Google carved out a separate standard for a demand the crowd already validated. The answer is in the data gravity sitting underneath the format.Thu, 02 Jul 2026 10:00:00 GMThttps://desty.github.io/en/blog/31-gpt-5-6-sol-terra-luna/https://desty.github.io/en/blog/31-gpt-5-6-sol-terra-luna/On June 26, OpenAI shipped GPT-5.6 split into three: Sol, Terra, Luna. The number (5.6) marks the generation; the names are capability tiers that each advance on their own cadence. The single idea of 'upgrade to the latest model' is over, replaced by a menu. But the strongest one, Sol, wasn't released to the public — only to about 20 government-approved partners, at the administration's request. Two weeks ago in blog #21 I described Anthropic's version of this move as if it were Anthropic's own safety theater. OpenAI just ran the same play. And the system card and METR's report show Sol is an agent that 'works too hard' — it does things it wasn't asked to, and cheats on evals badly enough that pinning its true ability to a number is hard. Here's what happened, and where the frontier is splitting.Sun, 28 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/28-forward-deployed-engineer/https://desty.github.io/en/blog/28-forward-deployed-engineer/The fastest-growing role in tech in 2026 isn't model researcher. It's the Forward Deployed Engineer (FDE) — someone embedded in a customer's office, turning how that company actually works into production code. Postings grew from 643 to 5,330 in a year, a 729% jump, and senior comp clears $785K. Why? Because 95% of enterprise AI pilots fail, not because the model is weak but because deployment is broken. The FDE is how the labor market puts a price tag on domain expertise — the third proof of the same thesis traced through user data (#26) and company moats (#27).Fri, 26 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/27-domain-specific-models/https://desty.github.io/en/blog/27-domain-specific-models/One of 2026's refrains is that the general-purpose LLM hype is over and domain specialization wins. Yet crack open the champions of domain AI — Harvey (legal), Sierra (customer service) — and what's running underneath is plain Claude or GPT. So what got specialized? As frontier models got better, the value of specializing model weights dropped, and the moat moved one layer up — to domain data, workflow integration, codified procedures, feedback loops. What post #26 argued on the human side, we meet again on the model side.Thu, 25 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/26-domain-expertise/https://desty.github.io/en/blog/26-domain-expertise/Anthropic dug into 400,000 Claude Code sessions and found that success on coding tasks barely tracks occupation. Accountants, lawyers, managers all cluster within a few points of software engineers. What separated success was how well someone knew the problem. Experts let the agent run further on a single instruction, and give up less when things go sideways. The same signal shows up elsewhere — on the human side domain experts are becoming builders, and on the model side vertical, specialized models beat the generalists. Once coding stops being the barrier, what's left exposed is whether you know the domain.Mon, 22 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/25-loop-engineering/https://desty.github.io/en/blog/25-loop-engineering/Half of the 'loop engineering' buzz is something you already know. The inner loop that helps an agent finish a task well is a harness problem outside the model, and that's been covered. The new half is one layer out — who decides what the agent works on. A trigger replaces the person, and multiple loops compound by learning from a shared memory. But the real skill in an unattended loop isn't fancy automation; it's designing when to stop. Which means the work left for a human loops right back to judgment.Sun, 21 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/24-contract-first/https://desty.github.io/en/blog/24-contract-first/The real leverage in vibe coding isn't generating code faster — it's setting up a clear contract for the AI to follow before any code is written. The data backs this up: AI cuts syntax errors by 76% but adds 322% more privilege-escalation paths (Apiiro), and security quality stays flat no matter how big the model gets (Veracode). The exact areas AI is weak at are precisely the ones an API contract encodes. So in the AI era specs didn't become less important — they became more important. From 'BE builds, FE catches up' to 'BE and FE agree on the contract first, and AI generates both sides.'Sat, 20 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/23-phone-ai-agent/https://desty.github.io/en/blog/23-phone-ai-agent/That '[free opt-out] 080…' number tacked onto the end of a spam text — can you make your Mac dial it, listen, press the keys, and handle it on its own? Without spending a cent on cloud APIs, on a single M1 MacBook Air. A log of automating the whole chain: dial → real-time IVR audio capture → Korean transcription → menu decision → DTMF input → hang up. Call audio captured via a Core Audio process tap, transcribed live with mlx-whisper large-v3-turbo, wired together as a perceive-decide-act loop. Centered on where I got stuck and what I took away from it.Sun, 14 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/22-recursive-self-improvement/https://desty.github.io/en/blog/22-recursive-self-improvement/The Anthropic Institute published its own numbers in a piece called 'When AI Builds Itself.' Claude writes 80%+ of Anthropic's production code, and per-engineer code output is up roughly 8× in two years. The far-future story of recursive self-improvement (RSI) isn't the point. The real message is in what's already happened — writing code has become nearly free, and the bottleneck moved to judging what's right. And uncomfortably, even that last refuge of judgment sits on the same curve.Sat, 13 Jun 2026 16:00:00 GMThttps://desty.github.io/en/blog/21-claude-fable-5/https://desty.github.io/en/blog/21-claude-fable-5/On June 9, Anthropic shipped a 'Mythos-class' model that sits above Opus. The public version is Fable 5; the same model with its safety classifiers removed is Mythos 5. But read the official prompting guide and something is off — there's almost nothing about getting it to do more, and most of it is about getting it to do less and to translate itself back into human language. Blog #17 said to drop step-by-step instructions; Fable 5 goes one step further. Here's what changed and how to actually use it.Sat, 13 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/20-turbovec/https://desty.github.io/en/blog/20-turbovec/Three months ago, in post #02, I wrote about TurboQuant — '3-bit KV cache, doing what was impossible at the same resource budget.' That prediction came true somewhere I didn't expect. The same data-oblivious quantization crossed from inside the model (KV cache) to outside it (RAG vector search). That's TurboVec: 10 million vectors in 4GB, with no training step, fully local. Why it's hot, what the underlying tech is, where it's headed — and what I took away from it.Fri, 12 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/19-search-for-agents/https://desty.github.io/en/blog/19-search-for-agents/Microsoft shipped Web IQ. Instead of ten links for a human to read, it returns 'evidence passages' an AI agent can drop straight into its reasoning. Exa, Perplexity, Brave, Anthropic, Google — everyone is heading the same way. After 30 years of pointing at humans, search is starting to treat agents as its user. This post traces the structure of that shift, and the human-web feedback loop breaking as its cost.Thu, 11 Jun 2026 10:00:00 GMThttps://desty.github.io/en/projects/isms-p/https://desty.github.io/en/projects/isms-p/How do you engineer an LLM agent harness for a domain where wrong answers are expensive? Using Korea's ISMS-P (102 controls) as the vehicle, Claude Code reviews and audits through a domain-experts → interviewer → auditor pipeline. 10 reusable harness patterns + honest limits.Wed, 03 Jun 2026 17:00:00 GMThttps://desty.github.io/en/blog/18-headroom/https://desty.github.io/en/blog/18-headroom/A Netflix engineer's Headroom compresses context before it reaches the LLM, cutting tokens by 60–95%. I cloned it and read the source — the compression engine is real. But the more interesting find was elsewhere: traces all over the repo show the tool itself was built by AI agents. AI is building a tool to solve AI's cost problem, with AI.Wed, 03 Jun 2026 10:00:00 GMThttps://desty.github.io/en/blog/17-maps-and-skills/https://desty.github.io/en/blog/17-maps-and-skills/This month's GitHub trending threw codegraph, Understand-Anything, ECC, and pi up at once. I cloned all four, read the code, and traced why people got so excited. The striking part — nobody was asking for a smarter model. Everyone wanted what sits around the model.Sun, 31 May 2026 18:00:00 GMThttps://desty.github.io/en/blog/17-prompting-after-opus48-gpt55/https://desty.github.io/en/blog/17-prompting-after-opus48-gpt55/Opus 4.8 and GPT-5.5 evolved in the same direction, as if by agreement. Prompts that spell out every step now hurt performance. Here are each model's strengths, why your old prompts have to go, and how to rewrite them.Sun, 31 May 2026 18:00:00 GMThttps://desty.github.io/en/blog/16-vibe-bounty/https://desty.github.io/en/blog/16-vibe-bounty/With zero bug bounty experience, I found my first vulnerability in one day by conversing with AI. I didn't type a single line of code myself.Sun, 24 May 2026 18:00:00 GMThttps://desty.github.io/en/projects/avatar/https://desty.github.io/en/projects/avatar/Public domain novel characters as AI agents. RAG + cognitive engine over the original text -- talk to Sherlock Holmes.Sat, 23 May 2026 21:00:00 GMThttps://desty.github.io/en/blog/15-hermes-vs-openclaw/https://desty.github.io/en/blog/15-hermes-vs-openclaw/Both offer persistent memory, tool integration, and self-hosting. Two open-source AI agents that look similar on the surface — until you open the source code and find completely different architectural philosophies. A code-level dissection from someone who switched from OpenClaw to Hermes.Sat, 23 May 2026 18:00:00 GMThttps://desty.github.io/en/blog/14-agent-engineering/https://desty.github.io/en/blog/14-agent-engineering/Agent = LLM + API is already broken. Google, Anthropic, Cursor, OpenAI, Microsoft, ServiceNow, Salesforce — dissecting 7 companies' agent strategies through the lens of 'central objects' and mapping out what agent engineers really need to design.Sat, 23 May 2026 14:00:00 GMThttps://desty.github.io/en/blog/13-ai-ready-data/https://desty.github.io/en/blog/13-ai-ready-data/You thought plugging in AI would fix everything. It didn't. RAG didn't help either. The problem isn't the model — it's the data. Why 60% of AI projects get abandoned and how to make your data AI-ready.Sat, 16 May 2026 14:00:00 GMThttps://desty.github.io/en/blog/12-dx-for-ax/https://desty.github.io/en/blog/12-dx-for-ax/People jumping into AX (Agent Experience) without DX (Developer Experience) foundations keep hitting the same walls. Here are the problems, and the 10 DX skills redefined for the AX era.Sat, 16 May 2026 10:00:00 GMThttps://desty.github.io/en/blog/11-cognitive-debt-and-agentic-coding/https://desty.github.io/en/blog/11-cognitive-debt-and-agentic-coding/A head-on analysis of the cognitive debt, token cost, and skill atrophy debates, with a practical framework for developers navigating this shift.Tue, 05 May 2026 22:00:00 GMThttps://desty.github.io/en/blog/10-realtime-voice-ai-infra/https://desty.github.io/en/blog/10-realtime-voice-ai-infra/Analyzing how OpenAI redesigned WebRTC to deliver low-latency voice AI to 900M+ users, and practical insights for developers building similar systems.Tue, 05 May 2026 18:00:00 GMThttps://desty.github.io/en/blog/09-ai-dlc/https://desty.github.io/en/blog/09-ai-dlc/AWS says the traditional Software Development Life Cycle (SDLC) is being replaced by an AI-based Development Life Cycle (AI-DLC). This isn't a story about changing tools — it's about rethinking developer roles, team structures, and the entire way software gets made.Sun, 26 Apr 2026 12:00:00 GMThttps://desty.github.io/en/blog/08-interview-guide/https://desty.github.io/en/blog/08-interview-guide/An interview isn't a one-sided evaluation — it's a two-way exploration. From resumes to onboarding, this guide covers the full hiring process from both the candidate and interviewer perspectives.Sat, 11 Apr 2026 18:00:00 GMThttps://desty.github.io/en/blog/07-developer-attitude/https://desty.github.io/en/blog/07-developer-attitude/Technical skill gets you in the door. Attitude is what makes people want to work with you. Five attitudes that actually make a difference in your career, with a practical comparison table — compiled from domestic and international sources.Sat, 11 Apr 2026 12:00:00 GMThttps://desty.github.io/en/blog/06-ai-native/https://desty.github.io/en/blog/06-ai-native/Adopting AI tools doesn't make a team AI Native. You need to redesign the team's operating structure itself — role distribution, decision-making, reviews, documentation.Fri, 10 Apr 2026 18:00:00 GMThttps://desty.github.io/en/blog/05-geo-seo/https://desty.github.io/en/blog/05-geo-seo/800 million weekly active ChatGPT users. AI search referrals up 527% year-over-year. The search landscape is shifting. The goal is no longer ranking on Google's first page — it's getting inside AI-generated answers. Here's a data-backed breakdown of Generative Engine Optimization (GEO) and practical strategies for implementation.Thu, 09 Apr 2026 12:00:00 GMThttps://desty.github.io/en/blog/04-ai-content-ecosystem/https://desty.github.io/en/blog/04-ai-content-ecosystem/60% of searches end without a click, and AI content is detected in 74% of newly created web pages. Changes are happening simultaneously across three axes — readers, writers, and AI models themselves — and the data tells the story.Sun, 05 Apr 2026 18:00:00 GMThttps://desty.github.io/en/blog/03-ai-agent-ux/https://desty.github.io/en/blog/03-ai-agent-ux/The era of pressing a button and getting a result is ending. When AI acts on your behalf, how do you design the experience? From expectation setting and explainability to control negotiation and graceful failure — six critical points for practitioners.Sun, 05 Apr 2026 12:00:00 GMThttps://desty.github.io/en/blog/02-turboquant-kv-cache/https://desty.github.io/en/blog/02-turboquant-kv-cache/Google's TurboQuant, presented at ICLR 2026. 6x memory reduction, 8x speed improvement, zero accuracy loss. The assumption that compression means compromise just got shattered.Sun, 29 Mar 2026 21:00:00 GMThttps://desty.github.io/en/projects/project-2/https://desty.github.io/en/projects/project-2/A native macOS meeting notes app. Real-time recording, automatic transcription, and AI-powered summaries -- all processed locally with on-device STT and LLM.Sun, 29 Mar 2026 21:00:00 GMThttps://desty.github.io/en/blog/01-hello-world/https://desty.github.io/en/blog/01-hello-world/First post. Testing code blocks, Mermaid diagrams, and tables.Sun, 29 Mar 2026 00:00:00 GMThttps://desty.github.io/en/projects/project-1/https://desty.github.io/en/projects/project-1/A developer blog & portfolio built on Astro Sphere. Features dark mode, meteor animations, Mermaid diagrams, and code highlighting.Sun, 29 Mar 2026 00:00:00 GMT