Governed Agent Systems

This programme studies how agentic AI systems can be made governable in practice. Rather than granting an agent broad autonomy, we route all consequential actions through one execution boundary that enforces policy, requires human approval for irreversible operations, records evidence, and supports rollback via compensating actions. The work is applied and iterative; it is not a validated production standard.

Agentic systems that can call tools and take real actions are difficult to trust: autonomy and accountability are in tension. Most deployments either over-restrict the agent (limiting usefulness) or under-govern it (risking irreversible harm).

• Can a single execution boundary enforce policy without materially reducing task success?
• Which classes of action genuinely require human approval, and how is that boundary drawn reliably?
• What evidence must be captured for an action to be auditable and reversible after the fact?

• Prototype a policy-enforcing execution boundary with a fixed tool interface.
• Instrument every action with evidence capture and a compensating-action registry.
• Run scenario-based evaluations comparing governed vs ungoverned task completion.

• Observed (internal, non-validated): routing actions through one boundary makes policy changes centrally auditable.
• Observed: drawing the irreversible-action boundary is domain-specific and remains an open design question.

• Formalise the irreversible-action classification.
• Externalise a reproducible scenario suite.

This programme is technical and does not involve community data.