TRACE
Evidence of what an agent did, under which constraints, in a form another party can check.
TRACE is work on verifiable execution for AI agents. The point is not a smarter agent. It is a record of execution that another party can inspect, replay, or verify without having to trust the operator.
Agent systems can call tools, touch private data, and change production state. Most logs are narrative, not evidence. If you cannot inspect what happened, you cannot put the agent in a real workflow.
I am working on the evidence layer: structured traces, verification of execution claims, and the developer infrastructure needed to make those records usable. The work sits at the intersection of agent runtime, cryptographic verification, and the same SDK instincts that shaped o1js.
An agent run produces a trace of actions, tool calls, and constraints. Verification checks that the recorded execution matches a policy or claim. The useful output is not a dashboard. It is a record another system can refuse.
Choices that mattered
Treat traces as evidence, not telemetry. A log that cannot be checked is not enough.
Keep the record inspectable by engineers, not only by the model that produced it.
Reuse verification ideas from zero-knowledge and attested execution instead of inventing a new trust model from slogans.
What came of it
Design and implementation work on verifiable execution records for agent workflows.
Related public infrastructure for isolated coding-agent execution in Code Airlock.
Positions agent systems as infrastructure that can be reviewed, constrained, and verified rather than trusted by default.