Every agent leaves a trail.
Trace, investigate and replay individual runs with the context needed for corporate accountability.
Know exactly what happened.
For every run, understand the full context behind it: what was decided, with which configuration and under which rule, captured the moment it happened rather than pieced together later.
Which agent ran?
Which version was active?
Which prompt was used?
Model
Which model was involved?
Policies
Which policies were applied?
Tools
Which resources were accessed?
Cost
How much did the run consume?
Result
What happened?
Replay the run, not the investigation.
Deterministic execution means a decision can be reproduced identically months later, not approximated from logs. The state at every step is the real one, with a checkpoint committed before the next step. When the regulator asks what happened, the answer is the run itself.
The record keeps the version, the prompt and the policies that were active in that run.
Time travel through a run.
Exact replay, step by step
The steps sit on a bar, the state of each one is the real one, and from there you can replay, fork or resume.
The violation names the rule
The audit line carries a rule id, not a prompt paragraph.
The interruption is the four-eyes check
The run pauses awaiting human approval and survives until someone decides.
Evidence by construction.
Hash-chained audit, per tenant
Append-only, enforced in the database and exportable to the SIEM. Tampering is detectable, not just discouraged.
Split retention
Forgetting a data subject's content leaves every metric, cost line and audit record standing.
Debug a run like a program
A breakpoint on a node, an operator pause across the cluster, time travel and fork.
Cost as a durable ledger
Real tokens per run, written to a ledger that survives restarts.