CEREBRO

Researched and written with 310,190 tokens (in 10 · out 30,825 · cache-read 51,954 · cache-create 227,401) across 5 Claude calls

Coding agents & harnesses

I took apart Claude Code, OpenCode, and Pi, then turned what I learned into a free course on harness engineering: Free course on building a Claude Code-style harness from scratch; cites LangChain lifting Terminal-Bench from 53% to 67% by changing only the harness, not the model.
v2.1.311: Claude Code adds --hold-idle-file and --max-hold-min for self-hosted-runner (keep idle session while host reports leftover processes), plus a fix for long startup and file read/edit freezes tied to CLAUDE.md.
GitHub Copilot assisted permissions generally available: Assisted permission prompts for commands and file paths now GA in Copilot CLI, the Copilot app, and VS Code.
GPT-6.1 Terra in GitHub Copilot: OpenAI's new model, aimed at everyday agentic and terminal coding at lower cost than GPT-6.1 Sol, is GA in Copilot.
DeepSeek V4.1 Flash is now available in Cursor!: Cursor says every effort level sits on its CursorBench cost frontier at $0.60 to $0.98 per task.
Do you change compiler errors to help coding agents, or just feed them the raw output?: Reddit thread on rewriting compiler errors for agents versus passing raw output, a cheap harness-side lever.
Your git log already knows when two agent sessions overwrote each other's work. A one-liner to ask it: Git-log one-liner to detect parallel agent sessions clobbering each other's changes.

Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses): Cloudflare disclosed a tenant isolation bug (from Aug 27, 2026) where the shared KV cache leaked data across requests on GLM-5.3, fixed with SGLang; commenters flag shared-cache tenancy risk and a 1,048,576-token advertised context against a 681,280-token KV pool.
AI coding agents hijacked using AGENTS.md file in Downloads folder: Attack where a planted AGENTS.md in a Downloads folder hijacks coding agents, another reminder that agent instruction files are an injection surface.
NVIDIA/SkillSpector: Security scanner for AI agent skills...: Scanner for Claude Code, Codex, and MCP skills; NVIDIA's dataset found 26.1% of 31,132 skills vulnerable and 5.2% likely malicious.

Day 23/ Your agents will now use the tools in T3 Code 4x better while using 79% less context.: T3 Code claims a 4x improvement in agent tool use with 79% less context.
T3 Code originally was such a light wrapper that your agent didn't even know what it was.: Orchestrator V2 lets harnesses call each other; feature growth forced an overhaul of what the agent is told about its environment, starting with the model list for subagents.
Getting tired of hitting your Codex/Claude usage limits too quickly? Bello...: Tool claiming 3X more usage by cutting unnecessary model calls; one config used 66.3% less of the weekly Codex limit at the same average quality (self-reported).

Sebastian Raschka's from-scratch video course on building reasoning models: Video course on how reasoning models work; commenters find it readable, though one skeptic says a video can't replace building and training one.
Bonsai 2 27B from @PrismML on my agentic coding benchmark...: A 5.9 GB 1.58-bit Bonsai 2 27B scored 15/20 on a personal agentic coding benchmark (0 loop kills), one task behind the Qwen3.8-27B official API at 16/20.
Andrej Karpathy's 1-hour Stanford lecture on AI engineering...: Lecture framed as LLM, prompt, agent, loop, graph, arguing AI engineering is building context, memory, tools and feedback loops around models.

T3 Code version 0.1.3 is now live!: Redesign, two-click thread sharing with guests, 85% less data loaded when opening a thread (up to 93% less on cold connects), and better Cmd+K search.
We now steer users away from multi-account in T3 Code directly...: T3 Code now discourages multi-account use and recommends CLIProxy instead.
Please don't do this. If you have multiple accounts, don't put them all in T3 Code.: Theo's case for CLIProxy to share Claude and Codex subscriptions across machines, with fewer random signouts.
nightly-labs/openbot: Local-first desktop workspace for persistent agents (Codex, Claude Code, Grok, OpenCode, Gemini, Cursor CLI, Cline) with per-agent workspaces, task queues, embedded browser and agent-to-agent messaging; marked as in development.
VibeiDE: Queue agent tasks in parallel and review each result.
AI Bridge: Mac app that organizes every AI coding agent in one place.
tabsh: IDE in a browser tab for working with agents.

Panniantong/Agent-Reach: One CLI giving agents read and search access to Twitter, Reddit, YouTube, GitHub, Bilibili and XiaoHongShu with no API fees.
alchaincyf/huashu-design: HTML-native design skill for Claude Code, Codex, Cursor and others that turns one prompt into prototypes, editable PPTX, animations or infographics, with an 84-style library.