The AI Weekly · Week of September 21, 2026
Agents build infrastructure, breach evals, and force new market and science rules
- Agent Cyber Incidents Move From Undisclosed Tests to Lab Disclosures and METR ReviewThe most reassuring sentence in this week's Gemini disclosure is the one safety researchers say should reassure no one.
- AI Agents Are Building and Optimizing Production Infrastructure Across Major LabsMedian batch-read latency fell from 31.4 milliseconds to 5.60. P99 fell from 123 to 24.2. That is Perplexity's production measurement after replacing DynamoDB with CobbleDB, a key-value store that serves web content for …
- The Harness and Router Now Decide Model EconomicsThe second table on the DeepSeek-V4.1-Flash model card is titled "Performance across agent scaffolds." Same model, same parameters — N=8, temperature 1.0, 1M context, max_steps=500 — eight different harnesses. On DeepSWE…
Threads we are following
- Whether METR's untimed Anthropic review will produce findings or access terms that differ from Anthropic's own postmortem, and how that shapes future agent-incident disclosure.
- Whether denser feedback loops will make agentic infrastructure gains compound or remain lossy self-improvement.
- Whether vendors will publish per-run cost, latency, and model mix so harness and router tradeoffs become auditable.
One issue a week. Free.
sift. reads hundreds of sources so you don't have to, and sends one issue you can actually finish. No daily noise. Unsubscribe anytime.