When One Agent Isn’t Enough
Sometimes a single agent can’t hold the whole job. So you split the work across several agents — they run in parallel, or hand off to each other, each focused on its own piece.
The Common Patterns
Most multi-agent setups are one of three shapes — or a mix of them.
- Subagents — a main agent delegates a focused sub-task to a helper, then uses the result.
- Parallel agents — several agents work at once on independent pieces.
- Pipelines — the output of one agent feeds the next in a chain.
Why It Helps
Each agent keeps a smaller, cleaner context and can specialise in one kind of work. You get breadth across a big task without one bloated session trying to track everything at once.
It Isn’t Free
More agents means more moving parts. Use them only when the work truly splits into independent pieces.
- Coordination — someone has to hand work out and track it.
- Merging — separate results have to be stitched back together.
- Review — more output means more for you to check.
A Good Default
Start with one agent. Reach for many only when the tasks are genuinely independent, or when the work simply won’t fit in one context. If a single agent can do it cleanly, let it.
Build It
How to implement: take a task with independent parts — say, three separate API endpoints — and give each part to its own agent run. Then review each result and integrate them together.
- Weekly AI Tasks tracker — build the UI, the API, and the message parser as separate agent tasks, then integrate them.
- Personal brand site — draft each section in its own focused run so every context stays clean and grounded.
What you learned
Multi-agent work comes in three shapes — subagents, parallel agents, and pipelines. It buys breadth with cleaner contexts, but adds coordination, merging and review. Start with one; add more only when the work truly splits.