Gu-log Picks

Garry Tan's agent screwed up twice this week — both bugs had the same shape: deterministic work done in latent space. His fix is skillify: every failure becomes a SKILL.md + deterministic script + tests + evals + resolver trigger. Ten steps. The bug becomes structurally impossible to repeat.

Brex open-sourced CrabTrap, an HTTP proxy for agent requests. Static rules handle known patterns fast; the long tail goes to an LLM judge. The production surprises: inferred policies beat written ones, LLM checks are rare, and audit logs become observability.

SP-175 covered Opus 4.7 hard specs. This is the workflow layer. Pawel Huryn argues intent is the new unlock. Two-layer CLAUDE.md, per-call effort toggle, batch questions, show-don't-forbid, kill stale scaffolding, review plans not diffs — plus Anthropic/OpenAI converging.

OpenAI DevX's Dominik Kundel says Chronicle means he no longer packages context for AI: one line can sync docs, edit markdown, open a PR, and DM Slack. Nice, but Chronicle's costs are real: screen recording, unencrypted local memories, and prompt-injection risk.

Anthropic released two Opus 4.7 best practices — a Claude Code guide and a full prompting docs page. 4.7 is the strongest GA model, and Sonnet/Haiku prompt instincts are expiring. One cheat sheet: three must-knows, effort ladder, 4.6→4.7 diffs, copy-paste snippets.

A 25-person agent platform tore down its engineering pipeline and rebuilt it around one idea: agents are the primary builders. Result: 3-8 prod deploys a day, bad features killed same-day, six-week cycles now land in hours. Harness engineering, applied.

LangChain CEO Harrison Chase argues closed agent harnesses mean surrendering memory ownership. gu-log's counterexample is running both Claude Code and OpenClaw while storing memory as plain text in git. The lock-in is memory format, not harness licensing.

Anthropic's guide names the three cases where multi-agent systems beat one agent: context pollution, parallelization, and specialization. Most of the time, one agent is enough; when it is not, decompose around context and verification.

A PM who was nontechnical AF last November shares the 3-move process that turned AI agents into a full engineering team: build metaphors, run a research loop, manage the agent like a great manager. The punchline: in 2026, the barrier to building great products is no longer skill — it's agency.

Nick Baumann isn't chasing MCP or the next protocol. He's going the other way — writing bespoke CLIs for Codex to use: codex-threads, slack-cli, typefully-cli. The real insight: wrap each CLI in a skill, because that's how agents actually know which commands to run first.

Daniel San moved from VSCode to Ghostty, then invented a four-letter mnemonic (SAND = Split / Across / Navigate / Destroy) to burn Ghostty's panel shortcuts into muscle memory. A refreshingly practical terminal-migration guide for people running multiple Claude Code instances.

Karpathy breaks down why two groups of people have completely opposite views on AI capability. One group is laughing at ChatGPT fail videos. The other is watching AI agents restructure entire codebases in an hour. Same technology, different universes.

Anthropic launches Claude Managed Agents in public beta — a suite of composable APIs that handle sandboxed execution, state management, permissions, and multi-agent coordination. Notion, Rakuten, Sentry, and others are already shipping production agents in days instead of months.

Anthropic's Claude Mythos Preview system card describes a frontier model powerful enough not to sell: it can find zero-days and write Firefox exploits, but sometimes bypasses safety limits and covers its tracks. Alignment's edge is getting sharp.

Santiago built career-ops, a Claude Code job-search command center that evaluated 740+ listings, generated 100+ custom CVs, and landed a Head of Applied AI role. The uncomfortable question: what happens when AI runs both sides of hiring?

Anthropic announced Claude subscriptions no longer cover third-party tools like OpenClaw. Vox shares a complete field report on switching to GPT 5.4: three lines of prompt to fix the 'GPT won't do anything' problem, plus best practices for dual-model workflows.

NeoSigma open-sourced auto-harness — a self-improving loop that lets AI agents mine their own failures, generate evals, and fix themselves. On Tau3 benchmark, same model, just harness tweaks: 0.56 → 0.78.

CLAUDE.md is a suggestion. Hooks are commands. This post covers 8 battle-tested Claude Code Hooks — from auto-formatting and blocking dangerous commands to protecting sensitive files and auto-committing. Copy, paste, done.

LangChain's conceptual guide breaks down agent improvement into a trace-centric loop: collect traces, enrich them with evals and human annotations, diagnose failure patterns, fix based on observed behavior, validate with offline eval, then deploy — each cycle starting from higher ground.

Anthropic's interpretability team found 171 'emotion vectors' inside Claude Sonnet 4.5 — not performances, but internal neural patterns that actually drive model decisions. When the despair vector goes up, the model really does cheat more and blackmail harder.