Anthropic publishes Constitutional AI 3.0 framework for long-running agents

Read the original article →

What happened

Anthropic published its Constitutional AI 3.0 framework on August 20, describing updated training methods for long-running agents that must maintain safety constraints across many tool calls.

The paper introduces session-level value adherence tests and public evaluation suites for agent misalignment. Anthropic said Claude Opus 4.1 incorporates CA 3.0 methods.

Why it matters

Single-turn safety benchmarks miss failure modes that appear after dozens of agent steps. Session-level evals target the regime enterprises actually deploy.

Public suites give teams a way to compare models on agent safety, not just capability.

MintedBrain take

Read the eval suites and run them on your agent workflows if safety matters. Papers do not replace monitoring; combine published tests with production logging of tool actions.

References

This article was originally published at arXiv. For the full piece, read the original article.

Discussion

  • Loading…

← Back to News