Frequently Asked Questions

Use multiple agents when a task has independent subproblems, separate tool or policy domains, useful independent critique, or context-window pressure that can be managed through artifacts. Start with one well-tooled agent first.

Sometimes. Anthropic reported a strong internal eval result for breadth-first research, but also reported much higher token use. Multi-agent systems work best when the task can exploit parallelism and the value justifies overhead.

A lead agent dynamically decomposes a task, delegates bounded subtasks to worker agents, and synthesizes their returned artifacts. It fits tasks where the needed subtasks cannot be fully predicted in advance.

Context fragmentation. Different agents may make incompatible implicit decisions if they cannot see the relevant state, traces, and artifacts produced by other agents.

Yes. A workflow follows a predefined path. A swarm is a looser group of agents coordinating through messages, shared state, or local rules. In practice, many systems called swarms are better described as manager-worker or pipeline workflows.

Use explicit artifacts, strong messages, clear handoff contracts, shared task state, traceable tool calls, and evals that judge both final output and process quality.

They can when independent critique, source checking, and evaluator rubrics catch errors. They can also amplify hallucinations if agents pass weak claims downstream without verification.

Build the simplest single-agent or workflow baseline, define evals, measure failure modes, then split only the part whose failure is clearly caused by prompt overload, tool overload, context limits, or independent parallel work.