Latest signal November 20, 2026 · afternoon edition
The newest agent gains come from fitting the harness rather than swapping the model, as Hugging Face trained small adapters to specific coding harnesses, Microsoft Foundry let browser agents run a whole click sequence as one script, and GitHub let team maintainers set default agent plugins for their teams.
01 / The wire
Recent briefings
-
November 20, 2026 · afternoon
The newest agent gains come from fitting the harness rather than swapping the model, as Hugging Face trained small adapters to specific coding harnesses, Microsoft Foundry let browser agents run a whole click sequence as one script, and GitHub let team maintainers set default agent plugins for their teams.
-
November 20, 2026 · morning
Agent governance is collapsing into single chokepoints (one identity per agent, one MCP URL per tool kit, one switch per blueprint), and a chokepoint only protects what cannot route around it.
-
November 19, 2026 · afternoon
On November 18 the control points of an agent moved out of the prompt and into the platform, as Microsoft put versioned tools behind one MCP endpoint and outbound traffic behind fail-closed rules while LangSmith tied coding-agent spend to the pull request it produced.
-
November 19, 2026 · morning
On November 18 agents moved from asking to staying connected, as GitHub put event streams into gh and added eight MCP connectors and Claude Code let MCP-backed policy hooks block on failure, so what happens when a check breaks is now the setting to audit.
-
November 18, 2026 · afternoon
On November 18 several research teams, Wiz among them crediting an AI harness, found critical NGINX bugs independently while OpenAI showed how it fences its own coding agent, and both stories point at the same gap, code that runs outside the fence.
-
November 18, 2026 · morning
On November 17 finding and doing got cheaper for agents, as a Codex-assisted audit surfaced a critical Java memory bug in an hour and Copilot gained a real browser, while the scarce work moved to checking what agents actually did.
-
November 17, 2026 · afternoon
On November 17 the unit of comparison for coding agents shifted from the model to the model plus its harness, as Runta measured harnesses changing cost per task severalfold with the model held fixed, Microsoft-owned GitHub added harness comparisons and Copilot Evals, and OpenAI and Microsoft pitched their new models on cost per task alongside capability.
-
November 17, 2026 · morning
On November 16 the rules around coding agents moved toward sources the work cannot rewrite, as the Claude Code team described review rules loaded from the base branch and GitHub let admins block standalone Copilot customizations, while the Optimizely compromise showed what a trusted but writable script costs.
02 / Under the surface
Latest analysis
-
Linger Puts a Personal AI Agent on Your PC. Whether Your Inbox Stays There Depends on One Setting.
Linger keeps your inbox on your machine only when its agents run on the local model, because its nightly wiki and memory jobs use whatever…
-
LangSmith Tool Gateway Makes Every AI Tool Call Pick an Identity: You or the Agent
Every agent tool call runs on someone's credentials, so decide per tool whether it acts as the user or as the agent, and put approval in…
-
Harness Adapters Turn a Local Model Into a Specialist for Your Coding Agent
A local model underperforms inside a coding agent partly because it was never trained on that agent's tools, prompts and control flow, and…
-
App Store Compliance Stops Your AI Agent From Shipping an App Store Rejection
App-store-compliance is most useful as a one-time scan you run by hand, because its agent hook only guards the submit commands an agent…
-
rust-anything Checks Whether Your AI Agent's Rust Rewrite Actually Works
Rust-anything makes the original program's recorded behavior the spec for an agent-written Rust port, so a pass means it matches every…
-
Microsoft Foundry's Crash Recovery Reruns Your AI Agent. Make Sure It Doesn't Email Twice.
Foundry hosted agents now recover from crashes by rerunning the handler from the top, which makes every side effect at-least-once, so the…
-
Claude Code Safety Hooks Fail Open by Default. One Setting Makes Them Actually Block.
A Claude Code policy hook that crashes, times out or cannot start lets the action through by default, so every hook you rely on as a gate…
-
Bitwarden Agent Access: You Don't Have to Give AI Agents Your Whole Vault Anymore
Bitwarden Agent Access narrows what an agent can reach to one approved vault item at a time, but only aac run keeps the secret out of the…
04 / Coverage map
Topics we track
Claude Code 126 Anthropic 53 OpenAI 44 GitHub Copilot 27 Codex 25 Agent Skills 19 Hugging Face 19 MCP 17 LangChain 15 Cloudflare 14 DeepSeek Harness 14 Model Context Protocol 13 Kimi K3 11 METR 11 Anthropic Frontier Red Team 8 GPT-5.6 Sol 8 Microsoft Foundry 8 Claude Code auto mode 7 GPT-6 Astra 7 MCP 2026-07-28 7 Claude Fable 5.1 6 Codex CLI 6 Cursor 6 GPT-5.6-Cyber 6