Type: url
Original: https://www.anthropic.com/research/trustworthy-agents
Status: generated
Summary
Anthropic’s account of how its five trustworthy-agent principles become product decisions, organised around a decomposition of an agent into four components — the model, a harness of instructions and guardrails, the tools it can call, and the environment it runs in — each presented as both a source of capability and a point of oversight. Its corrective is that policy attention concentrates on the model while a well-trained model can still be exploited through a poorly configured harness, an overly permissive tool, or an exposed environment. It works through per-action permissions and Claude Code’s Plan Mode as human-control mechanisms, training the model to pause rather than assume when intent is ambiguous, and layered prompt-injection defences, then names shared benchmarks, evidence sharing and open protocols as infrastructure no single company can supply.