AI work that checks itself: loops with real gates, graphs of agents, and where people approve.
An AI loop makes one piece of work correct by producing it, checking it with a test that can fail it, correcting what failed and trying again; a graph decides which pieces of work exist, which run in…
9 min read
Anthropic's October 22, 2026 guide argues that many company agents built as a router plus specialist subagents do better as one agent loop that loads skills and tools only when needed, and it ships an…
9 min read
Write the loop so that every piece of work passes through checks that can fail it without anyone watching, such as tests that must exit with code 0, a source line for every claim and a limit on which files…
6 min read
Start from the units of work and the checks each unit must pass, then add the nodes that connect them, which are a splitter that breaks the job into units, workers that run in parallel where units are…
8 min read
In research published on August 28, 2026, Anthropic had Claude agents run a loop of reading papers, proposing a training method, training a small model and scoring it, and on all ten alignment failures it…
9 min read
A check is real when it can fail the work on its own, with nobody watching, using evidence the worker cannot edit or argue with, such as a test that exits with an error, a claim with no matching source line,…
7 min read
Put the approval where a mistake would do real damage that is hard to undo, such as sending, paying, deleting, deploying or merging, not where the model sounds unsure.
8 min read
Sign up for the newsletter and receive the free guide: what changed in AI, one lesson, one prompt to try. Unsubscribe in one click.