Multi-Agent System Designer
Design a chief-of-staff AI agent with specialist agents: pick a pattern, roles, budgets and approvals, then export prompts, a handoff schema and SDK code.
The interactive tool loads here when JavaScript runs.
About this toolWhat is the Multi-Agent System Designer?
Describe a goal, choose how the work should flow, and pick the specialist agents you need. The designer writes a system prompt for a "chief of staff" orchestrator that plans the work, delegates tasks, tracks their status, checks results against acceptance criteria, merges them and escalates to a person, plus a focused prompt for every specialist.
It also produces the plumbing: a JSON Schema for the task and result messages agents pass to each other, a team diagram with its Mermaid source, Claude Code subagent files for .claude/agents, a Claude Agent SDK script and an OpenAI Agents SDK skeleton in Python. Four ready-made examples (a product page launch, a weekly sales report, a support inbox and a reviewed blog post) show each pattern. Everything is generated in your browser; no model is called.
- Four patterns: chief of staff, sequential pipeline, router, evaluator-optimizer
- Nine preset specialists plus custom agents with their own tools
- Budgets for turns, cost, time and revision rounds, and human approval checkpoints
- Chief-of-staff and specialist system prompts
- JSON Schema 2020-12 for task and result messages, with a valid example
- In-page diagram plus Mermaid source
- Claude Code subagent files, Claude Agent SDK script and OpenAI Agents SDK Python skeleton
- Zip download; nothing is sent to a server
Updated
How to use the Multi-Agent System Designer
- Load an example or type your goal: the outcome, who it is for and the deadline.
- Choose a pattern. Chief of staff suits open-ended work, pipeline suits repeatable stages, router suits inboxes, evaluator-optimizer suits work you can grade.
- Toggle preset specialists or add custom ones, and edit each one's description, responsibilities and tools.
- Set the turn, cost, time and revision limits and list the human approval checkpoints.
- Read the design review, then copy or download the prompts, schema, diagram and code, or download everything as a zip.
Worked examples
Product page launch
Input
Chief of staff with researcher, writer, editor and coder; 40 turns, $8, approvals on positioning and on the final page
Output
Chief prompt with a status board, four specialist prompts, four .claude/agents files with maxTurns 10
The per-agent turn limit is the team limit divided across specialists, rounded, with a floor of three.
Support inbox router
Input
Router with support, billing and bug reporter; approval on every refund
Output
OpenAI Agents SDK triage agent with handoffs=[support, billing, bug_reporter]
Routing hands the whole conversation to one specialist, so handoffs fit better than agents as tools.
Background and insight
Anthropic's engineering article "Building effective agents" (December 2024) separates workflows, where code fixes the path, from agents, where the model decides its next step. The patterns here map onto it: chief of staff is orchestrator-workers (a lead model breaks down the task and delegates subtasks it could not know in advance), sequential pipeline is prompt chaining, router is routing, and evaluator-optimizer pairs a generator with a grader in a loop. The same article advises starting with the simplest thing that works, which is why the designer warns when one agent would do.
Most multi-agent failures come from lost context and runaway loops rather than weak models. A specialist cannot see the orchestrator's conversation, so every task message has to carry the objective, context, inputs, deliverable and acceptance criteria in full. Budgets on turns, spend, wall-clock time and revision rounds stop a loop that is not converging, and explicit approval checkpoints keep a person in charge of anything public, costly or irreversible. The generated prompts write those rules down.
The exports follow each vendor's current documentation. Claude Code subagents are Markdown files in .claude/agents with YAML frontmatter where name and description are required and tools, model and maxTurns are optional. The Claude Agent SDK takes the same definitions in its agents option and delegates through the Agent tool. The OpenAI Agents SDK offers agent.as_tool for orchestrators that keep control and handoffs for routers that pass the conversation on.
Common use cases
Planning an agent project
Agree roles, limits and approval points with your team before anyone writes code.
Claude Code subagents
Generate .claude/agents files for a research, writing or coding team and drop them into a repository.
Prototyping with SDKs
Start from a working Claude Agent SDK or OpenAI Agents SDK skeleton instead of a blank file.
Operations automations
Design recurring jobs such as weekly reports or inbox triage with a person signing off the result.
Tips and limitations
- Try a single agent with good tools first. Add specialists only when one prompt gets too long or the subtasks need different tools or permissions.
- Write acceptance criteria the chief can check. "Every claim has a source" works; "high quality" does not.
- Keep teams small. Each handoff costs tokens and drops context; six or more specialists usually means roles should merge.
- Claude Code fails to launch a subagent whose tools list contains no valid tool names, so keep the Claude Code tools column to built-in names such as Read, Grep, Bash and WebSearch.
- The OpenAI Agents SDK has no dollar cap; the Python skeleton reads token usage from the run result so you can enforce one.
- Approval checkpoints only help if silence is never treated as a yes; the chief prompt says so explicitly.
Frequently asked questions
What does a chief-of-staff agent do?
It owns the outcome. It turns the goal into a definition of done, splits the work into tasks, delegates them to specialist agents with full context, tracks their status, checks results, merges them into one deliverable and escalates to a person when something is ambiguous, failing or risky.
When should I not use multiple agents?
When one agent with the right tools can do the job, when the subtasks share most of their context, or when you cannot yet measure quality. Extra agents add cost, latency and places for context to get lost.
What is the difference between handoffs and agents as tools?
With agents as tools the orchestrator calls a specialist, gets its answer back and stays in control. With a handoff the specialist takes over the conversation. Orchestrators use the first; routers use the second.
Where do Claude Code subagent files go?
In .claude/agents/ in the project (shared through Git) or ~/.claude/agents/ for all your projects. Each file is Markdown with YAML frontmatter, and the body becomes the subagent's system prompt.
Does the designer call an AI model?
No. It assembles prompts, schemas and code from your inputs in the browser. You run the generated code with your own API keys.
How do I stop agents looping forever?
Set a turn limit, a cost cap, a time limit and a maximum number of revision rounds, and tell the orchestrator to stop starting new work near a limit. The generated SDK code passes the turn limit and, for the Claude Agent SDK, a dollar budget.
Is there an in-depth guide to the Multi-Agent System Designer?
Yes. “The Chief of Staff Agent: Designing Multi-Agent Systems That Finish the Job” covers the background, how it works, common pitfalls and worked examples. Read the full guide.
Related Tools You May Like
Free, in-browser tools that pair well with this one.