Delegate to an agent only if it passes all five. RICE-A is the ADLC decision framework for deciding whether a task belongs to an AI agent or remains human-owned.
The delegation rule
RICE-A prevents unsafe delegation by forcing every task through five concrete checks before an agent is allowed to own execution.
The task is reversible, inspectable, constrained, low-cost if wrong, and assigned at the correct autonomy level. The agent can execute within approved guardrails.
A single failed criterion means the task carries risk the agent should not own independently. The agent may assist, but the human remains accountable.
The five criteria
Each criterion addresses a different failure mode in agentic development: reversibility, verifiability, ambiguity, blast radius, and autonomy fit.
Reversibility is the strongest safety property for agent-assigned work. If mistakes can be contained, rolled back, and corrected cheaply, the task is safer for agent execution.
Inspectable tasks produce outputs that can be checked by automated tests, linting, review, comparison, or evidence. If correctness cannot be inspected, delegation becomes guesswork.
Agents need boundaries. A constrained task has explicit acceptance criteria and measurable outcomes. Vague goals such as “make this better” are not safe agent-owned tasks.
Some tasks are reversible and testable but still dangerous because failure affects security, customers, reputation, or downstream teams. Low error cost keeps the blast radius bounded.
The first four criteria decide whether a task may be delegated. Autonomy decides how much freedom the agent receives: human-led, co-piloted, supervised, agent-led, or fully autonomous.
Apply at every gate
Do not apply RICE-A once and forget it. Re-apply it as scope, impact, data sensitivity, production exposure, and agent reliability change.
Outcomes
The agent executes within the approved autonomy level. Humans review checkpoint evidence rather than every micro-action.
The human performs or directly supervises the task. This is not a failure of the agent; it is the safety system working.
In practice
The same agent can own some tasks and not others. Ownership depends on the task’s reversibility, inspectability, constraints, error cost, and autonomy fit.
| Task | R | I | C | E | A | Owner |
|---|---|---|---|---|---|---|
| Write unit tests for a utility function | ✓ | ✓ | ✓ | ✓ | ✓ | Agent |
| Refactor a module to reduce duplication | ✓ | ✓ | ✓ | ✓ | ✓ | Agent |
| Generate inline code documentation | ✓ | ✓ | ✓ | ✓ | ✓ | Agent |
| Choose microservice architecture boundaries | ✗ | ✗ | ✗ | ✗ | ✗ | Human |
| Modify authentication token validation logic | ✓ | ✓ | ✓ | ✗ | ✗ | Human |
| Rename a public REST API endpoint | ✗ | ✓ | ✓ | ✗ | ✗ | Human |
| Fix a failing CI pipeline step | ✓ | ✓ | ✓ | ✓ | ✓ | Agent |
| Approve production release to 100% of users | ✗ | ✓ | ✓ | ✗ | ✗ | Human |
Key takeaway
RICE-A is not about distrusting agents. It is about earning safe autonomy with evidence.Passing the test shows that the task is bounded, verifiable, low-risk, and matched to the agent’s proven reliability. Failing the test keeps accountability with the human.
Continue the ADLC series
Keep the ADLC pages connected inside the DataKnobs ecosystem with previous, current, and next navigation.