An engineer and an auditor should read the same record.
AEGIS Trace is designed so that execution history, debugging, audit and evaluation all come from one ordered stream of events, instead of several systems that have to agree.
One event, two lenses
Same event. Two questions answered.
This is the denied INV-1048 payment from the simulated run. Neither reader needs to ask the other for a screenshot.
- Step
- decision → guard
- Model tier
- Reasoning
- Tokens
- 3,120 in / 540 out
- Attempts
- 1
- Error code
- policy_denied
- Next step
- Save as an evaluation case, then fix the agent's instructions
Tamper evidence
If someone edits history, it shows.
Each event includes the digest of the one before it. Alter an old event and every later link stops matching. Periodic anchors in write-once storage make it harder to rewrite the whole chain.
Tamper-evident means a change can be detected, not that a change is impossible.
- objective.createdFinance Controller (demo)aa05a9ae720cab30
- plan.validatedplan v1, 6 steps63460c7109dd7863
- action.proposedINV-1048, $2,750.00a0e270c5d937b7f3
- decision.deniedvendor_not_in_mastere2115feaf40a115c
- record.sealedexe_demo_A7K28ba3de17458bd165
Simulated record. The digest here is illustrative, not cryptographic.
Replay
Two kinds of replay. Neither repeats a side effect.
Deterministic replay
Rebuild the exact state from the record at any point, to see what the agent knew at that moment.
Counterfactual replay
Re-run a step with recorded tool results and a different model or prompt. Tools with side effects are never called again.
Example workflow
From incident to regression test.
Find
Filter by agent, status, reason code or time.
Inspect
Open the Mandate Map and see each step, decision and cost.
Save as a case
Turn this run into an evaluation case. Planned for Phase 2.
Fix
Publish a new agent version.
Compare
Run the suite on both versions, and block the release if it regresses.
Key capabilities
What it records.
| Capability | Release |
|---|---|
| Execution history and timeline | Phase 1 |
| Decisions, reasons and approvals | Phase 1 |
| Token, cost and latency per step | Phase 1 |
| Evaluation suites | Phase 2 |
| OpenTelemetry export | Phase 2 |
| Online evaluation and alerts | Phase 3 |
Technical explanation
Works alongside the tools you already use.
The Execution Record is the source of truth. Traces are a read model derived from it and can be rebuilt. Large payloads go to object storage with a digest in the event, and sensitive fields can be redacted before they are stored.
OpenTelemetry export is planned through an adapter, so traces can reach your existing observability tool. The GenAI semantic conventions, including agent spans, are still in Development status, so we expect that mapping to change.
Related capabilities
See a sealed record for yourself.
Complete the simulated run, then inspect every event.