Governance Levels

When deploying Agentic AI into production, "human-in-the-loop" is no longer a sufficient framing. Under emerging governance standards like the Singapore IMDA Model AI Governance Framework for Agentic AI (January 2026), organizations are expected to formally classify the autonomy level of their agents and provide corresponding human oversight.

Depute provides UI primitives that support implementation of each of these four governance dimensions.

---

The L1–L4 Autonomy Taxonomy

L1 — Human-Led

Agent proposes, human operates. The agent acts exclusively as an advisor. It generates data, intelligence, or suggestions, but the human must manually execute the final action in their own tooling.

• Depute Implementation: Use an Artifact Card or Tool Trace to display the agent's work cleanly. No complex execution barriers are needed because the agent lacks execution privileges in the host environment.

L2 — Collaborative

Agent and human work together. The agent can draft actions and stage tools, but requires active back-and-forth guidance and synchronous alignment.

• Depute Implementation: Use a Delegation Gate to smoothly manage the handoff when the human decides the agent is ready to take the reins. Wrap interactions in State Diff to show the human exactly what the agent drafted before it applies to the global state.

L3 — Supervised

Agent operates, human approves. The most critical boundary for high-stakes deployments. The agent is capable of running end-to-end but is cryptographically tethered to a human's explicit authorization before executing irreversible actions.

• Depute Implementation:
  • Wrap dangerous tools in an Approval Gate to halt execution immediately.
  • Surround the gate with an Automation Bias Alert to ensure the human doesn't blindly rubber-stamp the request out of fatigue.
  • Bind the final decision to a Decision Record to log an immutable audit trail of the human's approval logic for forensic review.

L4 — Autonomous

Agent operates, human observes. The agent runs asynchronously in the background across prolonged periods. The risk shifts from "preventing a single bad action" to "detecting cascading failures" across a massive, opaque blast radius.

• Depute Implementation:
  • Ensure users have persistent Run Controls to hard-kill the agent workflow at any moment from anywhere in the application.
  • Provide a Session Overview upon completion to cure "Execution Amnesia" and summarize the entire blast radius of the agent's autonomous run.
  • Use an Orchestrator View and Swarm Monitor to map the topology of multiple agents spinning up sub-agents autonomously.