A multi agent system divides work among multiple model driven workers. Context isolation gives each worker a bounded view of the task. The two often appear together, but a separate model call can provide isolation without a complex agent architecture.
Delegation is useful when tasks can be separated cleanly and their results can be checked. It is harder when workers need constant access to one another's changing assumptions or must edit the same state.
Identify work that can be delegated independently
Parallel document reviews are a plausible fit because each worker examines a different source and returns evidence relevant to one shared question. A coordinator then compares the findings.
A tightly coupled sequence may be a poor fit. If each step needs the previous step's result, extra agents may add communication without creating useful parallelism.
Anthropic's multi agent research system account describes benefits and costs in a particular research workload. Treat that experience as a design example, not evidence that every task benefits from multiple workers.
Give each worker a complete subtask contract
A delegation should specify the goal, available evidence, relevant constraints, permitted tools, expected result and resource limit.
Task: Find the termination-notice period in document LEASE-17.
Scope: Read only the supplied document.
Return:
- the notice period, if established
- the relevant clause and page
- a short supporting excerpt
- any exception or ambiguity affecting the answer
Do not decide whether the customer's current notice is valid.
The final sentence keeps a broader decision with the coordinator. If the worker also needs jurisdiction, contract version or customer dates, supply those details explicitly.
Require evidence at the handoff
The coordinator should receive source references and uncertainty alongside the finding. A confident conclusion without evidence is difficult to verify.
Validate returned structure and check source locations where feasible. Preserve disagreements between workers until they can be resolved against evidence. A majority vote may simply repeat a shared misunderstanding.
Do not overwrite one result with another because it arrived later. Use a result collection and an explicit synthesis step.
Control tools, shared state and budgets
A fresh context does not automatically provide separate permissions. Restrict each worker's tools and data access at the execution layer.
For shared writes, assign ownership or use concurrency controls. Two workers editing the same record can invalidate each other's assumptions even when both are individually competent.
Allocate work from a common run budget. The coordinator's usage and every worker's usage count toward the total. Cancellation should propagate so abandoned workers do not continue consuming resources or changing state.
Calculate latency and cost separately
Independent workers can reduce elapsed time because they run concurrently. Total token use may increase because instructions and source context are repeated.
The coordinator also spends time creating assignments, reading results and checking gaps. A worker that finishes late may determine the completion time for the whole task.
Compare cost per successful task, completion time, coverage and error rate against a single agent or fixed workflow baseline. Include orchestration failures in that comparison.
Test the coordination boundary
Test a worker that returns no evidence, a conflicting conclusion, a timeout and an incomplete result. The coordinator should distinguish these outcomes and avoid silently treating them as agreement.
A good delegation boundary leaves each worker enough context to answer its own question while making the returned evidence useful to the caller. If the handoff requires repeated clarification, reconsider the split before adding more agents.