Lesson Content

Multi-agent workflow design.

A multi-agent workflow has three layers:

1. Parent agent: Coordinates sub-agents, provides context, aggregates results
2. Sub-agents: Complete focused, scoped tasks and return results
3. Synthesis: Parent uses sub-agent results to produce final output

Design starts with decomposition: what are the distinct subtasks that benefit from focused context? Each distinct subtask is a candidate for a sub-agent.

Parallel vs. sequential sub-agents.

Parallel: Sub-agents that do not depend on each other's results can run simultaneously. A security sub-agent and a performance sub-agent analyzing the same codebase are independent – run them in parallel to save time.

Sequential: Sub-agents where the output of one informs the input of the next must run in order. A "discover failing tests" sub-agent must complete before a "fix failing tests" sub-agent can be scoped – it needs the list of failing tests.

Most multi-agent workflows have both patterns: some parallel sub-agents in the analysis phase, some sequential in the remediation phase.

The parent-aggregates pattern.

The parent agent should not attempt complex reasoning during aggregation – it should collect structured sub-agent outputs and combine them:

Phase 1 – Spawn and run parallel analysis sub-agents:
– Security sub-agent -> findings_security.md
– Performance sub-agent -> findings_performance.md
– Coverage sub-agent -> findings_coverage.md

Phase 2 – Aggregate:
– Combine all findings into a single report
– Deduplicate issues that appear in multiple analyses
– Prioritize by severity across all finding types

Phase 3 – Remediation (optional):
– Spawn remediation sub-agents for high-severity items
– Each remediation sub-agent receives the specific finding it addresses

Knowing when the workflow is too complex.

Multi-agent workflows have coordination overhead. If the workflow requires more than four to five sub-agents, complex dependency management, or multiple aggregation phases – consider whether the task can be decomposed differently, or whether the complexity warrants a more formal orchestration approach than ad-hoc Cowork mode sub-agents.