CEREBRO
machine-read, human-curated
Researched and written with 311,710 tokens (in 10 · out 28,954 · cache-read 64,753 · cache-create 217,993) across 5 Claude calls
Coding agents: releases & harnesses
- v2.1.327: Claude Code adds
claude mcp add-json <name> -to read server JSON from stdin, keeping tokens off the command line, plus a confirmation before/install-github-appchanges a repo. - v2.1.317: Claude Code now warns when
/scheduleattaches connectors you did not ask for, asks before sending your account email over the network from web/shell tools, and gainsonFailure: "block"for Agent SDK hook callbacks. - 0.170.0: Codex adds WebRTC negotiation for realtime voice in the TUI and auto-refreshed auth headers for OpenTelemetry HTTP exporters.
- What is the best harness for each model?: FrontierHarness v1.1 covers 27,280 runs and 83B tokens to test whether switching model should mean switching harness. HN pushback: the "64% of spend on failed runs" headline is editorialized, since "failed" only means missing pass criteria.
- obra/superpowers: Composable-skills methodology that makes the agent ask what you are building before it writes code. Worth a look for skill structure.
- affaan-m/ECC: Harness optimization bundle (skills, instincts, memory, security) spanning Claude Code, Codex, Opencode, and Cursor. Its README warns to install only from official channels.
- cursor/plugins: Cursor's official plugin spec and first-party plugins, each with a
.cursor-plugin/plugin.jsonmanifest. The plugin format is converging across agents. - Matt Pocock on constraining agents: Experiment called
/constrain: if too much choice hurts agents, allow React but force all state through Zustand/useReducer, enforced by harsh lint rules.
Agent infrastructure & safety
- External sandboxes for coding agents: Simon Willison explores running agent code in external sandboxes so a mistake or prompt injection from a local model cannot damage the host.
- Building the Agent Host Protocol: GitHub's open protocol for agent sessions that persist across editors, apps, phones, and cloud. Session portability is becoming a protocol layer.
GitHub Copilot updates
- GPT-6.2 Astra and GPT-6.2 Sol: Both models are now GA in Copilot, aimed at agentic coding and code review.
- MAI Code 2 and MAI Code 2 Flash: Microsoft's next-generation coding models land in Copilot.
- Copilot Fusion updates: Fusion gets new models, more Copilot surfaces, and results on two public Terminal-Bench leaderboards.
- Copilot Evals public preview: Lets you measure whether Copilot is still performing well as your prompts, models, skills, and tools change.
- Copilot in JetBrains, Ignite releases: Adds remote access, on-device voice dictation, browser-use controls, context usage shortcuts, and historical diffs.
Review & CI bottlenecks
- Harvard: AI coding agents shifted the bottleneck to code review: A working paper across 718 firms found about 30% more code, no measurable rise in finished work, and 49% longer review. It is not peer reviewed. HN pushback: the real bottleneck is up-front specification.
- Launch HN: StarSling: Agent-driven CI optimization that claims p95 test time dropping from 29m56s to 5m06s. HN comment: the metric that matters is deployable merges to main.
- Glass Page: A place for humans to review an agent's work.
Agentic SaaS & Mac tooling
- Hako: Puts every coding agent on your Mac in your pocket.
- Lasso: Draw on your Mac screen, then let AI agents see it.
CLI/TUI
- Unref: Rust Python dead-code finder on Ruff's parser, 20x faster than vulture, checking all of CPython in under a second. Built-in rules cover pytest, Pydantic, and SQLAlchemy.
Signals
- v2.1.327
- v2.1.317
- obra/superpowers: An agentic skills framework & software development methodology that works.
- affaan-m/ECC: The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Cl…
- External sandboxes for coding agents
- 0.170.0
- Launch HN: StarSling (YC P25) – Self-improving CI for agent-volume repos
- Building the Agent Host Protocol for sessions that continue across devices
- What is the best harness for each model?
- Harvard researchers find AI coding agents shifted the bottleneck to code review
- OpenAI’s GPT-6.2 Astra and GPT-6.2 Sol now available
- Hako
- Today's experiment: If too much choice is bad for agents, what if you could create an EXTREMELY constrained version of the tools you like?…
- cursor/plugins: Cursor plugin specification and official plugins
- GitHub Copilot Fusion updates: more availability, benchmarks, and models
- MAI Code 2 and MAI Code 2 Flash in GitHub Copilot
- GitHub Copilot in JetBrains, Ignite releases
- GitHub Copilot Evals is now in public preview
- Lasso
- Glass Page
- Show HN: Unref – A Python dead code finder built on Ruff's parser
Sources & citations
| Source | Fetched | In briefing |
|---|---|---|
| rss | 140 | 13 |
| hackernews | 30 | 4 |
| github | 36 | 3 |
| x | 216 | 1 |
| show_hn | 40 | 0 |
| github_search | 20 | 0 |
| swipe | 20 | 0 |
| yc-rfs | 11 | 0 |
| gmail | 7 | 0 |
| yc-launch | 4 | 0 |
| seed_urls | 1 | 0 |
| crackscan | 0 | 0 |
| ossinsight | 0 | 0 |
| 0 | 0 | |
| reddit_users | 0 | 0 |
| total | 525 | 21 |