Two Multi-Agent Patterns: Subagents vs Agent Team
Subagents (assembly line) vs Agent Team (newsroom) — architecture comparison, real benchmarks, and a decision tree for picking the right one.
A single agent hits three ceilings: the context window can’t hold complex tasks, one perspective introduces blind spots, and serial execution is slow. Multi-agent setups solve this. But Claude Code gives you two very different architectures to choose from.
Pattern One: Subagents (Assembly Line)
One Master Agent (the foreman) chains 6 sub-agents in sequence:
news-researcher → angle-picker → drafter → polisher → reviewer → formatterThe Master coordinates. It takes each sub-agent’s output and passes it to the next. All sub-agents share the main conversation’s context window.
Configuration lives in CLAUDE.md — you define each role’s responsibilities and handoff format. No extra installation needed. Claude Code supports this out of the box.
Pattern Two: Agent Team (Newsroom)
A Team Lead spawns 3 independent team members: researcher, writer, critic. Each member gets its own context window. They communicate directly via SendMessage, without routing through a coordinator.
You need to enable the experimental feature:
claude config set env.CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS 1The design has teeth. The researcher is obligated to flag data gaps proactively — things like “Bloomberg source is behind a paywall; used secondary source, cross-verified.” The critic has explicit rules against rubber-stamping. No “looks good to me” without substance. Writer and critic go back and forth for a maximum of 3 revision rounds.
Head-to-Head Benchmark
Same topic for both architectures: OpenAI’s $852B valuation vs Anthropic’s $900B valuation.
| Dimension | Subagents | Agent Team |
|---|---|---|
| Execution time | 14 minutes | 28 minutes (2x) |
| Sources cited | 5 | 8 |
| Narrative angle | Valuation metric shifts | Hidden truth in capital structure |
| Error handling | 3 post-hoc corrections | 3 pre-emptive risk flags |
| Token usage | Baseline | 3-4x |
With Subagents, the reviewer caught errors after the fact: a tense mistake, a unit conversion error ($9B written where it should have been $90B), and a missing source citation. The Master editor also flagged that an opening figure of $350B should have been $3,500B.
With Agent Team, there was no post-correction phase. The researcher flagged issues upfront: reliability of a paywalled Bloomberg source, the $900B valuation being a negotiation figure (not finalized), and inconsistent ARR disclosure methodologies between the two companies. These risks got handled before writing started.
The Agent Team also produced a sharper angle: “Of Amazon’s $35B investment, the conditions hinge on AGI/IPO — even the smartest capital won’t bet unconditionally.” That kind of insight came from multiple rounds of collision between researcher and critic.
Decision Tree
Use Subagents when:
- Batch production with standardized requirements (10 articles/day is feasible)
- Cost matters (3-4x lower token consumption)
- The workflow is fixed (same steps every time)
- Topics are straightforward (low fact density)
Use Agent Team when:
- Deep analysis with high fact density
- You need a novel angle (critic drives unique perspectives)
- The content is high-stakes (worth the token overhead)
- Factual errors carry real consequences
Hybrid approach (recommended): Route everyday content through Subagents. Send high-risk or high-importance pieces through Agent Team. The rule of thumb: “If getting facts wrong would hurt, use Agent Team.”
Configuration Templates
Subagents mode only needs role definitions and a workflow in CLAUDE.md, triggered with a standard prompt:
Write an article.
Topic: [your topic]
Execute the CLAUDE.md pipeline: 6 sub-agents run in sequence, final editorial review.
Output to outputs/ directory.Agent Team mode is similar but uses a different trigger:
Write an article using Agent Team mode.
Topic: [your topic]
Create researcher + writer + critic team. Let them coordinate via SendMessage.
Output to outputs/ directory.Customization
Swap roles: replace angle-picker with SEO-optimizer, add an image-selector. Change formats: adapt for Twitter threads, LinkedIn posts, newsletter editions. Tune parameters: word count (1000-1500), source count (3-5), revision rounds (2-3).
The architecture pattern stays constant. Specific roles and parameters flex to match your content type. One hard rule: keep team size under 5 members. Beyond that, communication overhead explodes and returns diminish fast.