Agent Profiler
Governing agents is a multidimensional challenge. This profiler illustrates how the dimensions interact, so stakeholders can reason about them together.
For illustration only. It is not an assessment of any agent: that rests on the agent's own record of judged runs, its undo and the guardrails actually in place.
- Reliability: how often it gets it right, as a lower bound over judged runs
- Blast radius: what happens when it doesn't, and whether it can be undone
- Business value: why it matters
- Governance thresholds: the bar to act alone, none for the irreversible
- Infrastructure: the guardrails in place
- Autonomy: the level of independence it earns
Built around the PAC Framework™: Potential, Accountability, Control.
What the Agent Profiler does
The Agent Profiler is a free, open-source tool for deciding how much autonomy an AI agent has earned. You describe a use case along six governance dimensions and the profiler translates that into an autonomy level, the thresholds that apply, and the infrastructure gaps that stand between the agent and the next level.
Reliability and blast radius set the risk: an agent that is right 95% of the time is fine when a mistake costs a retry, and unacceptable when it costs a customer or a compliance breach. Business value says whether the risk is worth taking at all. Governance thresholds mark where a human has to be in the loop. Infrastructure counts the guardrails that are actually there: sandboxing, audit trails, rollback, approval gates. Autonomy is the result, not the starting point.
Fourteen example cases, with capability scores drawn from public benchmarks such as SWE-bench, GPQA and BFCL, show what typical agents look like on the map. Every profile can be shared as a URL or exported as JSON, and you can set your own thresholds to match your organisation's risk appetite. It applies the PAC Framework, and it is the same tool participants use to prepare for the workshop.