Amey-Thakur Agent delegation protocol
The question is not what an agent can do but what it should. Capability is necessary and insufficient: the deciding factors are how reversible the action is, how visible a mistake would be, and who bears the consequence when it goes wrong.
Method
- Classify by reversibility first. Easily undone work delegates freely; irreversible action does not, regardless of how confident the agent seems.
- Keep the consequence with the human. Anything that spends money, binds the company, reaches a customer, or touches personal data stops at an approval (see agent-company-blueprint).
- Split rather than choosing all or nothing. Agents draft, gather, check, and prepare; humans approve, send, and commit. This split captures most of the value with little of the risk.
- Delegate the whole task, not fragments. Handing over an ambiguous fragment produces work that must be redone, which costs more than doing it (see agent-work-assignment).
- Define done before delegating. The acceptance criteria are what make the output checkable, and unspecified quality is the most common cause of disappointing agent work.
- Verify proportionally to consequence. Spot-check low-stakes output, review high-stakes output fully, and never accept unverified output that will be acted on irreversibly.
- Escalate on uncertainty by design. An agent that stops and asks when confidence is low is more valuable than one that always produces something (see agent-escalation-ladder).
Boundaries
This protocol governs delegation; it does not make an agent competent at the delegated task, which needs evaluation against real outcomes (see agent-eval-design). Regulated decisions, employment matters, and legal or medical judgements are not delegable regardless of capability. Accountability never delegates: the human remains answerable for what their agents do.