CEREBRO
machine-read, human-curated
Researched and written with 321,852 tokens (in 10 · out 38,171 · cache-read 63,963 · cache-create 219,708) across 5 Claude calls
Coding agents
- Claude Code v2.1.326 lets Code Mode run with agent teams enabled and warns on malformed
--allowedTools,--disallowedToolsor--toolsentries, so tool-allowlist typos stop failing silently. - Claude Code on desktop now supports iPhone Duo in the iOS simulator (public beta). Claude can fold and unfold the device to test each layout, which gives it a way to check foldable UI.
- I built TermPhone with Claude Code lets Claude drive an iPhone and Android device from the terminal, a mobile-control path outside the simulator.
- Fable 5.2 is the first model that consistently ignores unimportant feedback, per Theo. He says it ships on threads where Opus ran in circles, so it may be better at not over-reacting to review nits.
- VS Code retires the Local harness in February 2027. VS Code 1.146 deprecates the older Copilot agent-session harness, so migrate off it before then.
- GitHub Copilot weekly releases, November 9 adds GitHub Enterprise Server repos in the Copilot app, session triage in VS Code, and two new models.
- Control where GitHub Copilot customizations come from. Enterprise admins can require customizations to come from plugins or managed settings instead of user or project config.
- OpenCode Browse is an MCP server plus Chromium extension that gives agents your real browser (tabs, logins) and hands tabs back for 2FA or payment. HN commenters say Playwright MCP or chrome-devtools-mcp already cover this, and that Browse drops connections.
Token cost and context tricks
- DensePack claims a 45.0% to 52.8% cut in whole-session Claude Code cost with Opus 5.5, and 58.9% to 70.8% with a Sonnet subagent. These are self-reported numbers.
- Tokendoctor audits Claude Code overspend and claims up to 38% savings, from a 400K window and Sonnet subagents (context per request 412K to 219K). HN skeptics say the benchmark is thin, factory Claude Code often beats the tuned setup, and token tricks expire with each model release.
Sandboxing and agent security
- Simon Willison's external sandbox pattern runs the agent on your Mac but executes its code in a WebAssembly, QEMU or Apple Container sandbox, with demos for Claude Code, Codex and Pi.
- Each of his three sandbox servers has an optional HTTP/HTTPS allow-list proxy. A credential-replacement mechanism lets the agent make authenticated API calls without ever holding the secret.
- The index repository for the external sandbox prototypes was built with Prime Radiant, the lab behind Superpowers.
- Running coding agents without handing them my credentials introduces
lend, which lets remote agents use tools on your laptop with per-call approval. HN commenters already use proxy-based secret injection (Docker sandboxes, nono, smolvm), but note the proxy gets bypassed for Tailscale nodes. - mattpocock/sandcastle is a TypeScript library where one
sandcastle.run()call runs an agent in an isolated sandbox with a configurable branch strategy and merges its commits back. It is provider-agnostic, with native OS sandboxes such as Bubblewrap.
Harness design and agent quality
- Google Cloud's four-step harness for coding agents: context and memory, tools and workflows (MCP, ADK), sandboxes and prompt-injection defense, and evals (code, LLM judge, human).
- Matt Pocock on poteto's AIE Code talk: "nuance tends to lead to slop". Remove nuance from agent decisions by banning
useEffectand most explanatory comments, and by building APIs that hide complexity. - Theo on why he started T3 Code: he built it to lower his stress after every other agent tool disappointed him.
Agent skills and tooling
- kajisho5/ffmpeg-skill gives Claude Code, Cursor or Codex 42 local FFmpeg tools and an MCP server, with a machine-readable contract so the agent can check its own output. It needs no cloud or API keys.
- mvanhorn/last30days-skill is an agent skill that researches a topic across Reddit, X, YouTube, HN, Polymarket and the web, then writes a grounded summary ranked by engagement and real money.
- StateWeave lets you test your agent against working clones of the apps it uses, which suits regression-testing agent flows without touching production.
Vibe-coding and workflows
- A decade-experience engineer's vibe-coding steps is a practitioner workflow for side projects.
- QA skills that ship to 250,000+ users across 7 platforms replace manual flow testing with agent skills that act like a QA team.
- Building three Unreal Engine prototypes in parallel with an AI swarm covers the architecture, workflow and results from the Epic AI MegaJam.
- google-research/paper-orchestra is a multi-agent framework that turns ideas and experiment logs into submission-ready LaTeX papers, with literature synthesis and generated figures.
Signals
- mattpocock/sandcastle: Orchestrate sandboxed coding agents in TypeScript with sandcastle.run()
- DensePack cuts the price of whole Claude Code sessions by 45.0% to 52.8% with Opus 5.5, and 58.9% to 70.8% with a Sonnet subagent. It's lik…
- v2.1.326
- Claude Code on desktop now supports iPhone Duo in the iOS simulator (public beta). You can run your app on Duo next to your conversation an…
- Tokendoctor – the same Claude Code for up to 38% less
- Running coding agents without handing them my credentials
- How to build an AI harness around your coding agent in four steps: 1. Context and memory (sessions, state) 2. Tools and workflows (MCP, ADK…
- I've been exploring new ways to sandbox code written by coding agents - here's a description of the external sandbox pattern, where agents…
- I built TermPhone with Claude Code — now Claude can drive an iPhone (and Android) in the terminal
- GitHub Copilot weekly releases — November 9
- kajisho5/ffmpeg-skill: Local video editor for coding agents: 42 FFmpeg tools with a machine-readable contract and an MCP server, for Claude…
- VS Code retires the Local harness in February 2027
- mvanhorn/last30days-skill: AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesiz…
- I don’t think people know how hard I tried to make every other solution work before I gave up and started T3 Code. It wasn’t built to be po…
- .@poteto's talk at AIE Code is excellent. If I had to pick one still from that felt genuinely original, it would be this one: "nuance tends…
- Each of the three sandbox servers includes an optional HTTP/HTTPS allow-list proxy, so you can control exactly which hosts your agent's cod…
- Fable 5.2 is the first model that consistently ignores unimportant feedback. It’s incredible. Had at least a dozen instances where I used O…
- Control where GitHub Copilot customizations come from
- Building three Unreal Engine prototypes in parallel with an AI swarm: architecture, workflow, and results Epic Ai MegaJam
- google-research/paper-orchestra: PaperOrchestra is a multi-agent framework for automated AI research paper writing.
- StateWeave
- I worked on this project with Prime Radiant, the AI research lab behind Superpowers - here's our index repository for these external sandbo…
- I'm a software engineer with a decade of experience. I vibe code pretty much all of my side projects. Here are the steps I follow:
- I was spending more time manually going through flows than building, and still had bugs, until I created these skills. Now every release sh…
- OpenCode Browse
Sources & citations
| Source | Fetched | In briefing |
|---|---|---|
| x | 217 | 8 |
| rss | 140 | 5 |
| 50 | 5 | |
| github | 38 | 4 |
| hackernews | 30 | 3 |
| show_hn | 40 | 0 |
| github_search | 20 | 0 |
| swipe | 20 | 0 |
| yc-rfs | 11 | 0 |
| gmail | 9 | 0 |
| yc-launch | 4 | 0 |
| seed_urls | 1 | 0 |
| crackscan | 0 | 0 |
| ossinsight | 0 | 0 |
| reddit_users | 0 | 0 |
| total | 580 | 25 |