Single or Multiple Agents: Test the Value of Coordination

Reviewed and updated October 8, 2026.

Add multiple agents when independent work can run in parallel and the combined result is worth the coordination effort. Keep one agent or a fixed workflow when steps depend on shared state and must happen in order. Agent count is an implementation choice, not a measure of capability.

Compare a multi-agent design with a single-agent baseline on the same work. Measure useful outcomes, complete elapsed time, total model and tool use, failures and operator effort. Faster parallel execution can still use more total resources or produce a harder-to-review result.

Separate parallel work from dependent work

A request to compare independent supplier documents can be divided into bounded reviews. A request to amend one order, check its resulting state and notify the owner has dependencies; parallel agents may race on the same record.

Anthropic's account of its research system describes breadth-first research with a lead agent and workers, along with additional token use and coordination limits. It is evidence for that company's workload and implementation, not proof that multiple agents outperform one model at equal cost in every task.

An illustrative supplier comparison

Consider a comparison of three approved supplier proposals. This is a proposed evaluation, not an experiment that ALLTIPLY has run. A worker can review each proposal against the same criteria and return source references, missing information and findings. A coordinator checks comparability and produces the final table.

Define worker boundaries before starting: one assigned source, the agreed rubric, permitted tools, output schema and stopping rule. Keep workers read-only. The coordinator should verify important findings against the original documents rather than treating another agent's summary as authoritative evidence.

Compare this with one agent reviewing all three documents, and with a fixed extraction workflow. Include a conflicting definition, a missing proposal and one deliberately interrupted worker. Check whether the coordinator notices the gap or presents an incomplete comparison as finished.

Account for coordination failures

• Duplicate work: workers inspect the same source or repeat searches because assignments are vague.
• Lost context: a summary omits the exception that determines the decision.
• Incompatible outputs: workers interpret a field or rubric differently.
• State races: two agents mutate a shared record without a valid ordering rule.
• Unbounded continuation: the coordinator requests more work without a defined stopping condition.

A fixed workflow can be a useful alternative when subtasks and their ordering are known. Anthropic's effective-agent guidance distinguishes workflows from agents and recommends starting simply.

Make the comparison fair and useful

Declare the model choices, prompts, source access, task set and resource limits. If one design has a larger budget, report it rather than attributing the difference only to coordination. Repeat trials where variability matters and examine failure categories, not just an average score.

Choose the design that meets the decision's requirements with an acceptable operating burden. Adding workers is justified by demonstrated independent work and tested coordination, not by a universal penalty or benefit formula.

For a workflow architecture decision, see custom AI development. Request a project review with the task's independent parts, shared state and acceptable completion criteria.