AgentStateCrucible

Agents are non-deterministic, so a single successful run proves almost nothing. Exact-match assertions are too brittle for generated text, and pass/fail hides the thing that actually regressed.

Run the same scenario against multiple agents from the same starting state, capture every decision as a commit, and score the runs comparatively with a judge agent. Two agents can reach the same answer by very different routes; only trajectory comparison sees the difference.

• Same scenario, same starting state, same policy for every candidate agent
• Every decision a commit — intent, reasoning, confidence, alternatives, authority
• A pluggable LLM judge scoring runs side by side
• Sealed, Merkle-rooted epochs an auditor can verify without trusting the harness
• Measures correctness, reasoning quality, effect blast radius and cost

Scenario runner · judge scoring · run diffing · sealed audit epochs

• AgentStateDeveloper — validate behaviour and blast radius together

Source-available under BSL-1.1, converting to Apache 2.0 18 months after each version's first publication. Internal use — including in production — is unrestricted.