Agent identity
Which agent, version, configuration, or execution profile performed the work.
Provenance & Assurance Design
Make agent execution reconstructable.
Define the execution evidence that should be retained for the selected agent workflow — enough to understand later how the result was reached, without logging everything indiscriminately.
Which agent acted?
Under which process version?
Which context and rules applied?
Which tools and data were used?
Which outputs were produced?
Which checks ran?
Which human decisions occurred?
How was the final result reached?
Overview
Proportionate evidence — not maximal telemetry.
When an AI agent contributes to engineering work, teams may later need to understand which agent acted, under which process version, with which tools, under which controls, and with which human decisions in between.
Objective
Create a proportionate provenance model that supports review, debugging, governance, auditability, and high-assurance engineering where required.
Provenance dimensions
Seven dimensions the design can address.
Process context
Which operating-model and effective-process version applied.
Instructions and controls
Which skills, rules, hooks, DoD conditions, and authority constraints governed execution.
Tool use
Which tools, repositories, services, and interfaces were invoked.
Inputs and outputs
Which relevant artifacts were consumed, created, or modified.
Human interactions
Which reviews, approvals, overrides, or interventions occurred.
Execution history
The sequence of relevant actions and state changes required to understand the result.
How the service works
Five steps from assurance need to reviewable evidence.
Define assurance needs
Determine why provenance is needed and who may need to use it.
Identify material execution events
Avoid logging everything indiscriminately. Identify the evidence relevant to accountability, verification, troubleshooting, or audit.
Define retention and relationships
Specify how the evidence relates to the task, process, agent, tools, outputs, and approvals.
Design reviewability
Ensure the retained information can be interpreted by humans rather than existing only as machine telemetry.
Connect to assurance workflows
Define how provenance is used in reviews, investigations, assessments, or improvement.
Deliverables
- Provenance requirements
- Execution-evidence model
- Event and relationship definitions
- Retention recommendations
- Review model
- Assurance integration requirements
Benefits
- Reconstruct how a result was reached, not just what it was.
- Keep evidence proportionate to the assurance need.
- Make the record readable by a reviewer, not only by a machine.
- Give audits and investigations something to work from.
Tested by
Agent Provenance Review
A design says what should be retained. The review tests whether the real evidence is complete and usable.