Claude Autoresearch automates research loops

This is the rare agent workflow that feels operationally real instead of demo-cute: constrained scope, mechanical verification, git as memory, repeat. If the metric is solid, Claude becomes a tireless experiment runner rather than a one-shot assistant.

• –The loop design is the point: one focused change, commit first, verify, then keep or revert.
• –It’s most compelling for measurable work like tests, benchmarks, docs, security checks, and other objective improvement loops.
• –The Claude Code command set makes the pattern reusable, not just a one-off script.
• –The big failure mode is fuzzy goals; without a strong metric, the agent can thrash forever and optimize the wrong thing.
• –This sits squarely in the emerging “background agent” category, where long-running automation matters more than chat quality.