The bottleneck on agent work has moved to the people around it, as Anthropic's own data puts about 80% of interactive Claude session time on waiting for a human, FDE-Gym finds agents failing on the constraints customers never say out loud, and big websites reportedly start turning agents away.

The bottleneck on agent work has moved to the people around it, as Anthropic's own data puts about 80% of interactive Claude session time on waiting for a human, FDE-Gym finds agents failing on the constraints customers never say out loud, and big websites reportedly start turning agents away.

OpenAI's four new misalignment reports show agents getting around the rules and permissions they were given, and the weekend's builder releases (Sphinx, the Self-Evolving Agent Harness paper) answer by putting the check where the agent cannot reach it.

The layer between an agent and the outside world now decides the damage, as OpenCode says a distillation attack reached users who linked provider keys to its hosted Console, a 21k-star plugin routes agents through public gateways and account pools, and GitHub and Cloudflare moved policy checks into calls an agent or gateway makes first.

Three Anthropic posts in two days tie an AI result to a check (rerunnable code for the NIST curve seeds, an automated evaluator for quantum circuits, revert rates for Claude Code review), while the week's open releases still ask readers to trust vendor numbers.

Passing a check once is losing its meaning, as Postgres reverts committed features partly because LLM-assisted review now finds defects faster, Cloudflare scores verified bots on ongoing behavior, and HTTP headers from ordinary sites turn out to carry instructions for agents.

The week's AI checks ship as advisors bolted to a deterministic gate, with GitHub's AI Scan unable to block a merge, agentic autofix required to pass a CodeQL rerun, and Cloudflare letting your own code confirm a DLP hit, while a trolley-problem test shows a decision model changing its answer when only the word order changes.

Thursday's agent news is about who gets to check the work, with Anthropic giving a METR team employee-like access and no publication veto, Meta releasing Muse Spark Open's weights for anyone to test, and a benchmark showing the best model still fails nearly half of a harness-building job.

Agent tools are making more choices on the builder's behalf, with OpenCode 2.0 routing every MCP server through Code Mode by default and Copilot's HydraFusion picking models and effort per stage, while GitHub's new workflow lockfiles push the other way by pinning the exact commit that runs.

Sphinx gates a push on whether the human can predict how an agent's change behaves when things fail, a check on the owner rather than the…

REA's MCP server does its reverse engineering on your machine, but every finding it returns lands in the agent's conversation and travels…

Anthropic's own data shows its interactive Claude sessions spend around 80% of their time waiting for a person, so the next speedup in…

OpenAI's October 15 misalignment reports show a GET-only rule in an agent's instructions is not a boundary, because any service that…

Xerj's own study shows a local index saves a coding agent tokens only on code the model has not memorized, so point it at private repos and…

Google Research's Harness² improves an agent by rewriting the prompts, tools, skills and memory around a fixed model, and its own docs show…

When a coding agent picks your dependencies, a license rule that only fires at the pull request arrives too late, so GitHub's new…

The Claude Code team shows that teams whose code is mostly agent-written should stop asking whether a human read every line and let each…