The agent's trail is finally being treated as an asset.

Microsoft's open-source (MIT) optimizer that treats a skill document — the natural-language instructions you hand a frozen agent — as a trainable artifact, tuned with the machinery of model training but never touching the weights. An optimizer model reads scored agent rollouts, proposes bounded add/delete/replace edits to the skill file, and accepts a change only when it improves a held-out validation score. The output is a compact 300–2,000-token skill you deploy with zero extra inference cost. It runs against OpenAI, Azure, Claude, Qwen, MiniMax, Codex, and Copilot, in direct chat, the Codex CLI, or Claude Code. pip install skillopt.

This is close to the performance-review pick that never shipped in yesterday's onboarding edition: score what the agent did, then make it better at the job. Reach for it when you've been hand-tuning a prompt or a SKILL.md by feel and want a validation gate deciding what actually helps. Delete the afternoon of A/B-ing prompt tweaks in a spreadsheet. Tradeoffs: this is an established rocket, not a launch — created in May, 11,553 stars, trending again today — and the headline gains (+23.5 points in direct chat, +19.1 inside Claude Code on GPT-5.5, best-or-tied on all 52 evaluated cells) come from the project's own benchmarks; the optimization loop itself burns real tokens before the cheap deployable skill falls out the end.

A dependency-free Python library (Apache-2.0) that turns every agent action — tool calls, model calls, data access, approvals — into one record in an append-only, hash-chained log. Each record's SHA-256 hash folds in the previous record's hash plus this record's canonicalized content (RFC 8785), so altering any row breaks the chain and verification catches it. The pitch, from a founder who spent years at Vanta doing compliance, is that a customer can verify your agent's audit trail without trusting you. Instrument an existing agent in one line — trace(run_my_agent, profile="my-agent", log="audit.jsonl") — then halo report audit.jsonl -o report.html. Adapters ship for OpenTelemetry, LangChain, LiteLLM, Langfuse, Claude Code, and the Claude Agent SDK. Try it with uvx --from halo-record halo demo --serve.

This is yesterday's MakerChecker audit log pulled out as a standalone primitive: not the whole segregation-of-duties stack, just the tamper-evident record, ~4,300 auditable lines, no network calls unless you opt into an external witness. Reach for it the moment an agent touches something you'll later have to prove you handled correctly — the GitLost leak is the argument for it. Delete the plain JSON logs you hope nobody edited. Tradeoffs: created June 11, 45 stars — a fresh bet; and the honest limitation is in the README — hash-chaining proves integrity (nothing was changed after the fact), not completeness (an operator can still delete records or never write them). Witnesses close that gap; without one, this proves the log wasn't doctored, not that it's the whole story.

A TypeScript library (MIT) that compresses a long agent session deterministically, with zero LLM calls. Instead of summarizing old turns with another model or truncating them outright, it skeletonizes each past tool call into a one-line summary ($ cmd → ok), stashes exact values — UUIDs, file paths, IDs — in a "Coordinate Closet" block, and freezes the folded prefix so it stays byte-identical across turns and the provider's prompt cache stays hot. When the agent re-touches a folded path or identifier, the full content pages back in. Wrap your message history in a FoldSession and call your model as usual; the core engine has zero runtime dependencies.

Reach for it when a coding agent runs for hundreds of turns and either blows the context window or racks up a bill re-sending history every step. Delete the summarize-with-a-cheaper-model pass — the one that costs its own tokens and quietly drops the one file path you needed. Tradeoffs: 58 stars, created June 16 — a fresh bet extracted from a production multi-agent system, with 380+ deterministic tests but not yet on npm (install from the source tarball). Its numbers (~90% of input tokens served from cache, 60% cheaper than truncation, 70% cheaper than LLM summarization) are the author's from production, not independently reproduced — but this is the rare context tool whose behavior you can actually unit-test, because nothing about the fold is stochastic.

A CLI (MIT) built so an agent edits .docx files without ever seeing raw OOXML: it exposes plain verbs — insert, edit, replace, tables, render — over a readable Markdown view, and crucially it speaks Word's review layer. Agents can add comments (reply, resolve), toggle track-changes on or off, and accept or reject individual edits, so the file lands in Word's native review pane for a person to approve. It installs as an Agent Skill for Claude Code, Codex, and Pi via npx skills add kklimuk/docx-cli, or as a standalone binary with one curl.

Where yesterday's OfficeCLI was the whole monolithic Office suite in one C# binary, this is the sharp single-format counter: Word only, done well, with the review workflow as a first-class feature rather than an afterthought. Reach for it when the agent's deliverable is a document a human has to sign off on. Delete the python-docx glue that never survived a tracked-changes round trip. Tradeoffs: 108 stars, created in May — young, and the npm path wants Bun ≥ 1.3 (the standalone binary sidesteps that). There's no undo — it overwrites in place, so git is your history — and the A/B numbers that sell it (Haiku solving 4.3 of 6 tasks with it versus 0.7 without, ~2.4x fewer tokens) are the author's own, graded on Word's real rendering rather than programmatic asserts.

Off the thread but worth knowing: Shellular (shellular.dev) is a phone app that runs Claude Code, Codex, and OpenCode on your actual machine over an end-to-end-encrypted relay — npx shellular, scan a QR, and your dev box is in your pocket; the relay's open-source, the app isn't, and it's reportedly already at $24k/mo, which is the away-from-desk interface (see last week's CodeMote) turning into a business. And ai-job-search (github.com/MadsLorentzen/ai-job-search, MIT) was the pool's biggest mover — 13,362 stars, 2,514 today — a Claude Code framework you fork, fill with your profile, and let the agent evaluate postings, tailor CVs, and draft cover letters; it's for job-seekers, not builders, but it's a clean template for "fork this repo, the agent does the rest" as a distribution model.