Anthropic
Across this corpus, Anthropic is both a model developer and a publisher of first-party experiments about agentic coding, interpretability, and tool safety. Its reports make useful design evidence, but claims about capability, benchmarks, and deployed controls should be read as vendor-reported unless independently corroborated.
Recurring contributions
- Its long-running-agent work treats planning, generation, evaluation, structured handoffs, and browser-driven checking as separable harness roles; its C-compiler case study adds container isolation, task locks, and high-quality tests as practical coordination controls. [src] [src]
- Its safety material spans mechanistic interventions such as the Assistant Axis and operational controls such as Claude Code auto mode’s classifier-gated permissions. [src] [src]