Multi-Agent Patterns

One agent with good context handles most work. Multiple agents win when a task has separable roles (research vs. implement vs. review), needs parallel exploration, or has grown past what one context window can hold. Reach for these patterns then — not before.

When NOT to go multi-agent

Be honest about this first. Multi-agent setups cost more tokens, are harder to debug, and amplify bad handoffs. Skip them when:

• A single agent with a better AGENTS.md would fix the actual problem
• The task is sequential anyway (no parallelism to exploit)
• You haven't yet nailed a single-agent workflow for the task

The patterns below earn their complexity. Start with Pattern 2 — it's the workhorse.

Pattern 1: Researcher → Writer → Editor

The classic content pipeline.

• Researcher: gathers sources, verifies claims, produces research.md
• Writer: turns research into a structured draft — without doing new research
• Editor: tightens voice, checks facts against research.md, improves flow

Each role gets a fresh context window and a narrow charter. The separation is the point: a writer who can't research won't pad the draft with unverified claims.

Pattern 2: Planner → Implementer → Reviewer

The workhorse for code.

• Planner: explores the codebase, breaks the feature into tasks, writes plan.md with files-to-touch and risks
• Implementer(s): execute the plan exactly; deviations require an explicit note
• Reviewer: runs the advanced-code-reviewer skill against the diff; approves or sends back

Why it works: planning and implementing in one pass is where hallucinated architecture comes from. A written plan reviewed before implementation catches wrong directions when they cost nothing.

In Claude Code you can run this with subagents (each gets its own clean context); elsewhere, run sequential sessions handing off via files.

Pattern 3: Parallel Explorers → Synthesizer

For unknown territory — "why is this service slow?", "evaluate these three libraries".

• 2–4 Explorer agents investigate different angles in parallel, each writing findings to its own file
• One Synthesizer reads all findings and produces the final answer, noting where explorers disagreed

Parallel exploration is the one thing multi-agent does that a single agent fundamentally can't.

Handoffs: where multi-agent lives or dies

Every handoff should be a file with a fixed format, not a vibe. The minimum that works:

## HANDOFF — <from role> to <to role>

**Goal**: one sentence
**Done so far**: bullets, with file paths
**Artifacts**: research.md, plan.md, the diff…
**Open questions**: each with an owner
**Confidence**: high / medium / low — and why

Rules that keep it honest:

• Narrow charters — "You only write tests. Never edit implementation files." Agents respect hard scope limits far better than implied ones.
• Confidence is mandatory — a low-confidence handoff routes to a human instead of compounding downstream.
• Log every handoff to a scratch/ directory so humans can audit the chain later.

Copy the full templates: Planner → Coder handoff and the 4-agent team definition.