Coding agent harnesses and orchestration

• T3 Code adds Pi, delegate_task and an MCP server: one PR brings Pi support, auto-resume when rate limits reset, an MCP to create, message and wait on threads, and agents that spawn child agents on any provider, plus an ACP registry, OpenCode 2 and the official Cursor SDK.
• T3 Code cross-harness context handoff: lets you move a session's context between Claude, Codex, Pi and custom harnesses and back, which cuts provider lock-in for multi-agent workflows.
• T3 Code "Hide threads while working": a beta option that hides running threads so they don't take up space in your head.
• Theo on running 18 threads: threads surface only when they need attention, an interface pattern for scaling parallel agents without clutter.
• mattpocock/sandcastle: a TypeScript sandcastle.run() runs coding agents in Docker, Podman or Vercel sandboxes on their own branches and merges the commits back.
• Matt Pocock on GitHub Actions software factories: issues become agent tasks, labels trigger Actions that open PRs, and labels applied by Actions form loops, giving a near-free sandbox for public repos.
• Pi 1.0, Pi Durable, and AIE NYC: Pi 1.0 and Pi Durable ship under Earendil, which Pi has joined, and Pi is now being treated as a serious OpenClaw-class harness.
• Theo on running 6 Claude and 3 Codex subs: a video on stacking many subscriptions, with a warning that it may get you banned.

Benchmarks, nerfs and model reliability

• Theo on prompt-outcome volatility: Claude Code updates, provider changes, cache-hit shifts and seeds can all change results within minutes, so public benchmarks are meaningless without controlled access.
• Theo disputes the Opus 4.6 "nerf" claim: says the April claim rested on only 6 of 30 benchmark tests, with 2 failures treated as proof of a nerf.
• BridgeMind defends NerfBench: the other side of the dispute, with a thread explaining how NerfBench works.
• Theo's misinformation video list: upcoming topics include model nerfs, whether token prices and TPS matter, the real impact of bad harnesses, and "compaction = deletion".

Tokens, speed and model guidance

• JuliusBrussee/caveman: a skill and proxy that cuts output tokens about 65% by forcing terse caveman-style responses, backed by an Adobe Research paper reporting 1.4 to 3x lower cost.
• What if AI worked at 1,000,000 tokens per second?: at that speed, deciding which draft to trust becomes the bottleneck, and commenters say latency, not throughput, is what limits real-time applications.
• A model guide for the GPT-6 family: OpenAI's guidance on choosing GPT-6 models, tuning reasoning effort, writing prompts and skills, coordinating tools and preparing workflows for production.
• Coachwhip: an open-source one-line install that runs the 80B Qwen3-Coder-Next fully offline on a 16 GB Apple silicon Mac.

Agent tools, MCP and skills

• Claude Code skill that filters AI review comments: checks CodeRabbit comments against the code, removing 34% of the noise while keeping 93% of real bugs.
• Pony: gives Claude control of an Android phone via MCP or your own Anthropic key.
• Claude as game master over MCP: Claude Code built an RPG world engine, and the same engine now exposes MCP so Claude runs the game.
• Agent-Reach: one CLI that lets an agent read and search Twitter, Reddit, YouTube, GitHub, Bilibili and XiaoHongShu with no API fees.
• WeftCut: an open-source video editor designed to be driven by an AI agent.
• JarvisCore: builds agents as zero-trust peers in a mesh network.

Agentic SaaS

• Cue by Manus: personal agents with persistent identity that handle tasks end to end.
• Famulor: an agent that answers calls and handles follow-ups on WhatsApp.