An agent result without its path is difficult to inspect and expensive to trust.

We encountered that problem in two different instruments. Rooklight, the harness behind our Pi agent work, coordinates tool-using sessions. AlphaSwarm-SOL coordinates specialised roles for Solidity security analysis and uses a behavioural security knowledge graph.

Output is not the record#

A security result cannot be reduced to a confident sentence. The useful material is the route to that result: the assumptions, the attempted attack path, the rejected branch and the verifier’s objection.

The same rule applies outside security. If an automation changes a file or recommends an action, the operator needs enough state to reproduce what happened. A trace is not decorative observability. It is part of the product surface.

Roles create disagreement on purpose#

AlphaSwarm-SOL uses attacker, defender and verifier roles. The point is not to imitate a team meeting. The point is to stop one generation from serving as its own evidence.

That mechanism still has a hard limit. Multiple agents can share the same bad premise. A behavioural graph can preserve an analysis without making it correct. Human security review remains required, and the instrument cannot prove that a contract has no vulnerabilities.

Documentation is the interface#

A harness earns trust by making state inspectable and failures recoverable. That requires more than displaying model prose. It means recording tool calls, constraints and the point where execution stopped.

We publish the limit beside the mechanism because the two statements belong together: the system shows more of its work, and showing more work is not the same as guaranteeing the answer.