Tim Tyler of Gravitee presents agent governance as an organizational responsibility in this 18-minute AI Engineer talk. His workforce analogy adds ownership and oversight to existing software controls.

Assignment: name the agent and its sponsor

  • Give each agent a unique identity in a shared registry, instead of making all activity look like its user’s account.
  • Record its team and the context it should exchange with that team.
  • Name a human sponsor who owns the agent’s purpose, spending and response when something goes wrong.
  • Tyler compares explicit ownership to adults acknowledging who is watching children at a pool.

Alignment: separate the job from each request

  • The sponsor defines the lasting role: purpose, responsibilities, budget and success criteria.
  • The agent’s current mission describes the immediate task; it must remain within that role and its granted authority.
  • Grant only the data and tool access needed, with clear escalation boundaries.
  • In Tyler’s delegation example, only information allowed across the whole chain returns to the original requester.

Administration: check conduct and results separately

  • An agent can stay within its permissions and still fail its job.
  • Check hard rules directly, such as payment limits or allowed write destinations.
  • Judge less certain goals, such as correct invoice matching, using a scorecard and explicit confidence assumptions.
  • Monitor the model, incoming context and outgoing work continuously; “guardian” agents are Tyler’s proposed automated oversight layer.

Define the intervention before the incident

  • Specify when to correct behavior, reduce authority, escalate to a person, block an action or suspend the agent.
  • Gateways and authorization systems enforce restrictions; the sponsor revises the role and permissions as evidence accumulates.
  • Tyler’s confidence percentages are illustrative policy examples, with no demonstrated calibration or tested universal thresholds.
  • Caveat: this is a vendor-sponsored governance framework, not evidence that guardian agents reliably enforce every rule or detect every failure.

“Rules are checked, intents are judged.” — Tyler, 13:15