CEREBRO

Researched and written with 319,765 tokens (in 10 · out 39,358 · cache-read 66,075 · cache-create 214,322) across 5 Claude calls

Claude Code and Anthropic agent design

Anthropic's new guide (Oct 22) argues most agents should use fewer subagents (one main loop plus skills and tool search) and says Claude Code's system prompt was cut by over 80%.
How Anthropic's Dynamic Workflows Coordinate Claude Agents covers "flows", a new Claude Managed Agents feature announced Friday, from the AI Engineer NYC workshop.
Thariq Shihipar's Fable 5 talk (summarized in Chinese) frames "unhobbling": capabilities surface through use, and giving a model code execution unlocks tasks it can't do in its head, which Anthropic calls capability overhang.
MCP Tasks and Elicitations in Claude: Tasks let plugins run long jobs like deep research and report back when done.
Donating Agent Skills to the Agentic AI Foundation: Anthropic is moving the Skills standard under the Linux Foundation's AAIF, alongside MCP, so skills are now vendor-neutral.
ClaudeDevs on the donation says 70+ platforms have integrated Skills since December 2025, including ChatGPT, Cursor, Copilot and VS Code.
Boris Cherny on the donation notes Skills began as a Claude Code experiment.
Claude Code v2.1.309 fixes sandbox escapes: shell startup files ran outside the sandbox before each Bash command, and macOS denyWrite globs failed when the folder was reached through a symlink.
Claude Code v2.1.310 brings the Monitor tool and plugin monitors to Bedrock, Vertex and Foundry (with long foreground sleeps now blocked there), and fixes gateways that reject the effort-2025-11-24 beta value.

Codex 0.164.0 adds ten numbered /presets slots for saving reusable prompts and slash commands, with a keyboard menu.
Recursive Flint on a 20-task Python bug-fix benchmark scored 13/20 on a single RTX 3060 at about 49 tok/s, using 2.5B active params at 3-bit. That's a useful data point for local agentic coding, since the 4-bit 35B MoEs next to it manage 15-16 tok/s.

Corgi is a terminal multiplexer that tracks each agent's state (working, needs you, finished, idle); HN commenters call it a v2 of herdr and ask whether it should be a herdr plugin.
NVIDIA OpenShell is a sandboxed runtime for autonomous agents that lets them read files, install packages and use credentials without unrestricted access.
Peras pauses Claude Code and Codex before they burn through your weekly limit.
Inyon lets you answer your coding agents from your phone.
Shared-room MCP server for Claude sessions lets multiple Claude sessions work together, and the author used it to build itself.

Serena gives Claude Code symbol-level find and edit, so it doesn't read whole files or grep around, which saves context.
mq is jq for Markdown: it extracts the smallest complete section, heading or code block, so an agent doesn't have to read the whole document.
Swipe: Search the work your AI agents already did features a local search engine over past coding-agent sessions, plus a codebase map and a private inbox for projects and agents.
Swipe: Claude in plain English rounds up a plugin that rewrites Claudish, a menu bar app that swaps accounts at the rate limit, local agent memory, and a skill that stops agents quitting early.

review-pr is a tight review contract that defines what doesn't count (speculative warnings, style nits, untouched code) so agent reviews stay actionable.
t3dotgg/tstack is Theo's daily skill set for Claude Code and Codex, including full-send, which takes a change to a merged PR unattended.
shot-scraper video records a Playwright video from a storyboard.yml, so coding agents can produce demo videos of their work.
Open skills directory offers one-command installs of reusable skills into Claude Code, Cursor, Codex and Gemini; the post lists ten worth installing.
motion2code is a new MIT skill that rebuilds a motion-design clip as inline SVG with one GSAP timeline and checks it against the source.