Run an Agent Team
Complex tasks overwhelm a single AI agent — the context gets muddy, quality drops, and it does everything serially. A small team of specialized agents, each with a focused role and clean context, can tackle it in parallel and check each other's work. This designs that team for your task: the roles, how they coordinate and hand off, what context each needs (and what to isolate), and the guardrails — turning "one agent doing everything" into a coordinated effort.
What This Skill Produces
- The task decomposition — the task broken into distinct agent roles, each with a focused responsibility (researcher, drafter, critic, integrator, etc.)
- A coordination pattern — whether agents run in parallel or sequence, how their outputs combine, and where the hand-offs are
- Context design — what each agent needs to know, and (crucially) what to keep isolated so one agent's context doesn't muddy another's (the key to why teams beat one agent)
- A review/quality step — a separate agent or pass to critique and integrate, so quality is checked, not assumed
- Guardrails — how to keep the team on track (clear objectives, defined outputs, a human checkpoint) and avoid runaway loops or drift
- A right-sized recommendation — including when a single agent is genuinely better (not everything needs a team)
Required Inputs
Ask for these if not provided:
- The task — the complex thing you want a team to tackle
- Your setup — the AI tool/framework you're using (Claude Code sub-agents, an agent framework, or manual multi-chat)
- The subtasks — the natural pieces, if you can see them
- Quality bar & stakes — how much the output matters (drives the review rigor)
- Constraints — cost, time, and how much human oversight you want
Framework: Decompose, Isolate, Coordinate, Review
- Check it needs a team. Not every task does — if it's simple or highly sequential with shared context, one agent is better. Use a team when parts are genuinely parallel or benefit from distinct, isolated perspectives.
- Decompose into roles. Break the task into focused responsibilities, each an agent — a researcher, a builder, a critic, an integrator — so each has one clear job.
- Isolate context deliberately. The power of a team is clean, separate context per agent — decide what each needs and what to keep apart, so perspectives stay distinct and context stays sharp.
- Choose the coordination pattern. Parallel (independent then combine), sequential (hand-offs), or a mix — and define exactly how outputs pass between agents and merge.
- Add a review pass. A separate critic/integrator step catches errors and combines the work — don't trust unreviewed parallel output.
- Guardrail it. Clear objectives, defined output formats, iteration limits, and a human checkpoint keep the team from drifting or looping.
Output Format
Agent team: task [x] · setup [y]
Needs a team? [yes — parts are parallel/benefit from isolation / no — one agent is better because Z].
Roles
| Agent |
Responsibility |
Context it needs / isolate |
| [researcher] |
|
|
| [builder] |
|
|
| [critic] |
|
|
| [integrator] |
|
|
Coordination: [parallel / sequential / mix] — outputs combine by [how].
Review pass: [critic/integrator checks & merges].
Guardrails: clear objectives · defined outputs · iteration limit · human checkpoint.
Quality Checks
Anti-Patterns
- Using a team for a task one agent handles better.
- Agents with muddy, shared context (loses the whole advantage).
- No review pass — trusting unchecked parallel output.
- Vague roles that overlap and conflict.
- Missing guardrails — runaway loops or drift with no human checkpoint.
Example Trigger Phrases
- "How do I use multiple AI agents to build this?"
- "Set up an agent team to research and write this report."
- "Orchestrate several agents for this complex task."
- "Should this be one agent or a team, and how do I structure it?"
- "Design a parallel agent workflow for this."
1---2name: run-an-agent-team3description: "Design a small team of AI agents to tackle a complex task in parallel — who does what, how they hand off, and how to keep them coordinated — instead of one overloaded agent doing everything serially. Use when asked how do I use multiple AI agents, set up an agent team, orchestrate agents for, or run agents in parallel. Produces a decomposition of the task into agent roles, a coordination pattern (parallel vs sequential, how outputs combine), the context each agent needs (and what to keep isolated), a review/quality step, and the guardrails to keep it from going off the rails — practical multi-agent design for real tasks."4disable-model-invocation: true # skill-optimizer5metadata:6 source: https://skillmd.com/skills/slipcast-dev/run-an-agent-team7---89# Run an Agent Team1011Complex tasks overwhelm a single AI agent — the context gets muddy, quality drops, and it does everything serially. A small team of specialized agents, each with a focused role and clean context, can tackle it in parallel and check each other's work. This designs that team for your task: the roles, how they coordinate and hand off, what context each needs (and what to isolate), and the guardrails — turning "one agent doing everything" into a coordinated effort.1213## What This Skill Produces1415- **The task decomposition** — the task broken into distinct agent roles, each with a focused responsibility (researcher, drafter, critic, integrator, etc.)16- **A coordination pattern** — whether agents run in parallel or sequence, how their outputs combine, and where the hand-offs are17- **Context design** — what each agent needs to know, and (crucially) what to keep *isolated* so one agent's context doesn't muddy another's (the key to why teams beat one agent)18- **A review/quality step** — a separate agent or pass to critique and integrate, so quality is checked, not assumed19- **Guardrails** — how to keep the team on track (clear objectives, defined outputs, a human checkpoint) and avoid runaway loops or drift20- **A right-sized recommendation** — including when a single agent is genuinely better (not everything needs a team)2122## Required Inputs2324Ask for these if not provided:25- **The task** — the complex thing you want a team to tackle26- **Your setup** — the AI tool/framework you're using (Claude Code sub-agents, an agent framework, or manual multi-chat)27- **The subtasks** — the natural pieces, if you can see them28- **Quality bar & stakes** — how much the output matters (drives the review rigor)29- **Constraints** — cost, time, and how much human oversight you want3031## Framework: Decompose, Isolate, Coordinate, Review32331. **Check it needs a team.** Not every task does — if it's simple or highly sequential with shared context, one agent is better. Use a team when parts are genuinely parallel or benefit from distinct, isolated perspectives.342. **Decompose into roles.** Break the task into focused responsibilities, each an agent — a researcher, a builder, a critic, an integrator — so each has one clear job.353. **Isolate context deliberately.** The power of a team is clean, separate context per agent — decide what each needs and what to keep apart, so perspectives stay distinct and context stays sharp.364. **Choose the coordination pattern.** Parallel (independent then combine), sequential (hand-offs), or a mix — and define exactly how outputs pass between agents and merge.375. **Add a review pass.** A separate critic/integrator step catches errors and combines the work — don't trust unreviewed parallel output.386. **Guardrail it.** Clear objectives, defined output formats, iteration limits, and a human checkpoint keep the team from drifting or looping.3940## Output Format4142### Agent team: task [x] · setup [y]4344**Needs a team?** [yes — parts are parallel/benefit from isolation / no — one agent is better because Z].45**Roles**46| Agent | Responsibility | Context it needs / isolate |47|---|---|---|48| [researcher] | | |49| [builder] | | |50| [critic] | | |51| [integrator] | | |5253**Coordination:** [parallel / sequential / mix] — outputs combine by [how].54**Review pass:** [critic/integrator checks & merges].55**Guardrails:** clear objectives · defined outputs · iteration limit · human checkpoint.5657## Quality Checks58- [ ] Checks whether a team is actually warranted (vs one agent)59- [ ] Decomposes into focused agent roles60- [ ] Deliberately designs isolated vs shared context (the key advantage)61- [ ] Defines the coordination pattern and how outputs combine62- [ ] Includes a review/integration pass63- [ ] Adds guardrails against drift and runaway loops6465## Anti-Patterns66- **Using a team** for a task one agent handles better.67- **Agents with muddy, shared context** (loses the whole advantage).68- **No review pass** — trusting unchecked parallel output.69- **Vague roles** that overlap and conflict.70- **Missing guardrails** — runaway loops or drift with no human checkpoint.7172## Example Trigger Phrases73- "How do I use multiple AI agents to build this?"74- "Set up an agent team to research and write this report."75- "Orchestrate several agents for this complex task."76- "Should this be one agent or a team, and how do I structure it?"77- "Design a parallel agent workflow for this."