What you will learn
Supervisor patterns, Handoffs, Shared state, Parallelism, Merge conflicts.
Engineering the capability
Multiple agents can separate responsibilities, but they add handoffs, latency, and inconsistent assumptions. Define each role’s input, output, allowed tools, and stopping rule. A specialist should return structured findings with evidence, not an unbounded transcript for the next model to interpret.
Use shared durable state for facts that must agree. Avoid allowing two workers to overwrite the same field without conflict handling. Parallel work is appropriate for independent subtasks; dependent tasks must wait for valid prerequisites. A supervisor should reconcile incompatible findings and preserve uncertainty rather than blindly selecting the most confident response.
Compare against a single-agent baseline on the same tasks. More agents can increase token use without improving correctness. Measure handoff failures separately so you can see whether the decomposition is helping or creating new errors.
Worked case
A researcher finds a policy allowing refunds; a verifier finds a later restriction. Combining their prose by concatenation produces a contradictory answer. Require each finding to include scope and effective date, then have the merge step detect the conflict before drafting a customer response.
Put it into practice
Continue with the next lab: compare one agent with two specialists. Build the artifact, record the failure cases, and explain the tradeoff before moving on.