AI agent audit
A gap report, a risk map, and a ranked fix list for the operating layer around agents that already run — not a model bake-off.
Typical timeline: From a written brief to a delivered audit is about a week when access is ready. A sprint or retainer is a separate decision after you have the report.
This is the flagship. Production failures of AI agents come from a missing management layer, not from the model: unclear tasks, no safety gates, no monitoring, no one who owns the stop. We know the pattern because we run more than ten agents in daily production on our own work.
The audit looks at the six layers of the harness — including the management layer most checklists skip. When the harness is sound, the report says so and points at the real cause instead of inventing a model upgrade.
For whom
Teams whose agents work in the demo and surprise them in production
Operators who need a management layer (OODA, PDCA, TOC) rather than another prompt pack
Leads who want a part-time agent-operations function they cannot hire full-time
What you get
- Gap report of the running system against a harness and an operating cadence
- Risk map: where an agent can act unsafely, silently, or without an owner
- Ranked fix list you can run without us in the room
Formats
Harness Health Check
Diagnose. Fixed-scope audit. You get the report, the map, and the ranked fixes. This is the default first engagement.
Management OS Sprint
Install. A short engagement that puts in the gates, the monitoring cadence, and a playbook your team can run.
Fractional Advisory
Operate. A monthly retainer: review of the fleet, new agents brought in under the same bar, on-call when something breaks.
How it goes
- 01
Written brief: what the agents do, where they have surprised you, what you will not tolerate
- 02
Harness Health Check against task spec, gates, monitoring, escalation and ownership
- 03
Report: gap list, risk map, ranked fixes
- 04
You take the report and run
or we install the layer in a sprint — or we stay as fractional advisory
What we do not do
We do not start with a sales call. We do not replace your model vendor. We do not certify a fleet as safe because the dashboard is green. A green dashboard is a claim; the checker stays independent of the generator.