What changed after we gave support agents write access?
The model was not the hard part. Approval boundaries and recovery paths were.
By sourcetms
AI news for people who apply it.
A quiet place for consequential updates, firsthand context, and conversations with people doing the work.
Visit sourcetms01 / Briefing
The AI news feed
Editorial snapshot
- 01
GPT-5.6’s real story is efficiency, not another benchmark win
The useful shift is in the stack around the model: caching, inference, and agent harnesses are becoming the product advantage.
- 02
A first map of how AI is actually entering the economy
Google’s ATLAS data points toward augmentation before autonomy. Adoption is broad, but the work is still mostly shared with people.
- 03
OpenAI turns forward deployment into a product
Presence packages policies, evaluations, approved actions, and escalation. Production reliability is moving above the model layer.
- 04
Claude moves from a private chat into the shared room
Claude Tag treats the agent as a scoped team member. The important design problem becomes permissions, memory, and visible delegation.
02 / Community
Concept preview
Applied
A forum for what happens after the announcement. Real workflows, honest postmortems, and questions worth answering.
Built for practitioners, not spectators.
Representative discussion previewRead-only
A practical budget for parallel coding agents
Comparing throughput, review load, and the point where more agents stop helping.
The smallest useful internal AI workflow
One trigger, one narrow decision, one human checkpoint. It beat the ambitious version.