Signal, not noise: five primary-source reads from September 14 to 15, each with a two-line note on why it matters for builders.
Agentic coding is straining CI. Here is how we scaled test impact analysis at Anthropic
Anthropic engineering reports CI job volume up 25x in six months as output rose and test counts grew, with three stopgap fixes buying 70 days, then 29 days, then less than a day before a redesign around an in-memory journal plus stateless workers. Plan your own selection service for roughly an order of magnitude more load than today, and keep state out of the process so it can scale horizontally.
Read the full story at Claude by Anthropic (September 14, 2026).
Configure cost and quality in Copilot auto model selection
GitHub adds three auto tiers, efficiency, balance, and intelligence, so teams can state how auto should trade cost, quality, and response time per prompt across supported surfaces. Pick one tier for routine edits and a higher tier for hard tasks, then check bills against the model auto actually selected.
Read the full story at GitHub Changelog (September 14, 2026).
Give every teammate and agent the right level of access to your Workers
Cloudflare now scopes access down to a single Worker with four roles covering metadata inspection, code reading, editing without deletion, and full management, assignable to users or API tokens for agents. Hand each agent a token that can touch only its own Worker, and keep broader accounts out of everyday agent runs.
Read the full story at Cloudflare Blog (September 15, 2026).
Have it both ways: stay discoverable in search while disallowing AI training
Cloudflare introduces a training opt-out that keeps accountable mixed-use crawlers allowed for search while blocking training use, replacing the blunt block-everything tradeoff with synced robots preferences. Set the training preference once centrally instead of hand-editing robots files per operator, and review ad-page presets before the defaults apply.
Read the full story at Cloudflare Blog (September 15, 2026).
What If the Adaptation Were a Model? ShadowPEFT in PEFT library
Hugging Face documents ShadowPEFT as a first-class fine-tuning method in PEFT with different adapter geometry from LoRA but the same entry point and save-load path, merged to main for the next release. Try the development install on one adapter experiment first and compare directly against your LoRA baseline before standardizing.
Read the full story at Hugging Face Blog (September 15, 2026).