Researched and written with 318,132 tokens (in 10 · out 33,147 · cache-read 66,435 · cache-create 218,540) across 5 Claude calls

Coding agents: token cost and context

• Claude Code sends each background agent's transcript to the model: a gist shows a BUN_OPTIONS preload that disables the periodic agent_summary calls (tested on 2.1.257 to 2.1.312), and one team reports those calls took 16.4% of weekly token spend. A commenter says it's the agents-panel summary rather than the status line, and prompt caching may make the repeated reads cheap, so measure before patching.
• Context window is not memory: a context window doesn't persist across sessions, and compaction or session end can bury useful detail, so you must decide what to preserve, where to store it, and how to retrieve it.

Capability inflection

• The October inflection point for vibe game development: Simon Willison argues October 2026 is a second step change after November 2025, with Claude Opus 5.5 now building video games that are actually fun (his post, Boris Cherny's RT).

GitHub Copilot and Actions go agentic

• GitHub Agentic Workflows GA: GitHub Next's open source project lets you write Markdown automation that compiles to Actions YAML, for reasoning tasks like issue triage, CI failure analysis and doc updates (announcement on X).
• Copilot Fusion GA: you pick Fusion instead of a model, and Copilot routes and orchestrates models per task.
• Copilot Fusion local model selection: an experimental preview that lets the router choose between on-device and cloud models.
• Shared Copilot Automations: teams can co-author workflows that combine prompts, triggers and tools.
• GitHub Copilot app, Universe releases: the app now handles work spanning multiple repos, specs, designs and assets in one session.
• GitHub intelligence: agents get your team's preferences, repo conventions, skills, custom instructions and history as context.

Multi-machine agent harnesses (T3 Code)

• T3 Code multi-machine orchestration, internal milestone: Theo reports one thread spanning multiple models, harnesses, apps, OSes, machines and networks.
• T3 Code fleet control via a single MCP server: Theo spun up 8 threads on different machines from a Grok Bot.
• Remote control is over the internet, not computer use: the fleet stays controllable from a phone with the laptop off.
• T3 Code orchestrating across connected remotes: a user can say "build iOS/Mac stuff on my Linux boxes, run tests on my Mac remotes", and calls it a gamechanger.
• fx is now a built-in T3 Code provider: Vercel's open-source agent harness runs on web, desktop and mobile with your existing models and skills.

Governance, skills and enterprise

• Bernstein: an open-source governance layer where a deterministic Python scheduler, with no model in the coordination loop, runs agents in parallel, replays plans byte-identically, and enforces policy-as-code with a verifiable record.
• expo/skills: official agent skills for building, deploying, upgrading and debugging Expo and EAS apps, pointing agents at the real docs and CLIs as the source of truth.
• Claude Enterprise GA additions: smart reports, in-app surveys and pooled group budgets became generally available on Oct 28.

MCP, CLI and tooling

• levelsio's MCP directory submission: directories reject API-key-only auth, so he added OAuth login, and the MCP now works inside Claude and ChatGPT for city rankings and trip management.
• Use the CLI, not the chat UI, for AI coding: a Reddit nudge toward terminal agents.
• Yardlet: a Mac app that wraps the coding agents you already use.
• Zest: a hotkey (hold ⌥Space) that hands a dev chore to an AI agent.
• Chinese-language index of 41 practitioner posts: organised into 9 categories including Claude Code, harness orchestration, Skills and Grok Bot, useful as a bookmarkable reference.