Amey-Thakur Agent work assignment
Assignment fails in two directions: everything routed to the most capable worker until they are the bottleneck, or work spread evenly onto whoever is free regardless of fit. Good assignment considers both capability and current load, and refuses when neither works.
Method
- Describe the work before assigning it. The deliverable, the constraints, and what done looks like. Vague work returns vague output regardless of who does it.
- Match on capability, then on load. The best fit that has capacity beats the best fit that does not, because queued work delivers nothing.
- Give every worker an explicit capacity. Concurrent items or tokens or hours, stated. Without a limit, assignment continues until quality collapses (see agent-capacity-planning).
- Refuse rather than overload. When nothing has capacity, the correct output is a message saying so with the trade-off, not a silent queue.
- Assign to one owner, always. Shared ownership produces mutual assumption. Collaborators are named separately from the owner.
- Include the context the worker needs, and no more. Enough to act, without the whole company's state, which wastes context and leaks (see agent-context-isolation).
- Confirm acceptance before counting it as assigned. A worker that cannot do the work must say so immediately rather than at the deadline (see agent-handoff-protocol).
Boundaries
Assignment allocates work; it does not create capacity, and a system that is genuinely oversubscribed needs scope cut or resources added by a human. Assigning work to people carries fairness and employment implications and is a management responsibility, not an agent's (see agent-people-ops-desk). Capability estimates for agents are approximate and need real outcomes to calibrate.