Build a team of AI agents
You've built a single agent before. Now build a small team: one Supervisor that plans, and up to four Specialists that carry out the work, and watch it plan, run, pause for your approval, hit a snag, and recover. Still no code. Your team stays in this browser and is never sent anywhere.
Listen to this briefing
Building Resilient, Autonomous AI Teams
New to AI agents?
This tool builds on Build Your Own AI Agent, the beginner tool. If you haven't built a single agent before, start there first: it takes 5–10 minutes and explains the basics this tool assumes you already know: goals, tools, and rules.
What's different here
One agent. One goal. A flat, simulated reasoning trace. A single "ask first" rule.
A Supervisor + Specialist team. A visible plan before it runs. A live run with an approval checkpoint. A simulated failure and recovery. A team-wide governance record.
What you will do: 5 steps
- 1Start. Load a starter team, or build one from blank.
- 2Build your team. Name your Supervisor, then add Specialists with jobs and tools.
- 3Give it a task. Write a task, then see the plan your team makes before anything runs.
- 4Watch it run. Approve a checkpoint if one comes up. Watch it recover if a tool fails.
- 5Get your team's record. A full governance report for the whole team, ready to print or save as a PDF.
Your team is kept in this browser as you go, so you can leave and come back. Press Start over to clear it.
Step 1 of 5: Choose a starting point
Choose a starting point
Pick a starter team that matches a real scenario, or start blank and build your own. You can rename and rewire everything after.
Every real deployment starts from a known-good shape. These three templates are simplified versions of patterns used by real multi-agent platforms: a Supervisor that plans, and Specialists that each do one job well.
Build your team
Every team has one Supervisor and up to four Specialists. The Supervisor never does the work itself: it plans, assigns, and checks.
Your Supervisor
Plans the task, assigns each step to a Specialist, and merges the results.
Tools (pick at least one)
A team splits one big job into smaller jobs different agents are good at. Marking a Specialist as needing approval creates a checkpoint the plan can never skip: this is how real teams keep a person in the loop on risky actions.
Give it a task
Write the task below. Your Supervisor will read it and build a plan. Nothing runs until you approve the plan.
This is the plan your Supervisor has made. Nothing has run yet.
This is a practice tool that teaches the ideas behind multi-agent AI using simple, built-in rules. It does not call a real AI model. Everything you type stays in your browser.
Seeing the plan before it runs is the single biggest upgrade from a single agent. You can catch a wrong assignment or a missing checkpoint before anything happens.
Watch your team work
Each step lights up as it runs. If a checkpoint comes up, the team pauses for you.
Real teams fail: a tool times out, an approval takes a moment. This step shows the whole loop: plan, act, check, and (when needed) recover, live.
Your team's governance record
The same idea as the beginner tool's record, scaled up to a whole team and to the extra risks a team brings.
This record is simplified for learning. It cites real, published design patterns from the Agentic AI Pattern Catalogue: 61 patterns for autonomous agents, mapped to real regulatory standards including Singapore's IMDA Model Governance Framework, the OWASP Top 10 for Agentic Applications, and the US NIST AI Risk Management Framework. Finishing this tool does not mean a real agent team is certified or compliant. For a real organisation, use the full ISO 42001 AIMS Readiness Checklist and TRACE Agent Evaluation tools.
Four ideas this tool adds
You already know goal, tools, rules, and prompt from the beginner tool. A team adds four more ideas.
Supervisor
The agent that plans and assigns, but never does the work itself. It keeps the team on track.
Plan
The list of steps the Supervisor makes before anything runs, so you can check it first.
Checkpoint
A step marked to always pause for a person's approval before it runs.
Recovery
What the Supervisor does when a tool or step fails, instead of just stopping.
Where the four ideas appear in one team run
Schematic of a run in this tool with the Customer Refund Approval Chain starter team: Supervisor Ren, Policy Checker (marked as needing approval) and Customer Reply. The tool simulates a tool failure on the first Specialist step of every run.
The part of this tool I'm proudest of is the failure. Every team you build here hits a tool that doesn't respond, and has to recover mid-run instead of just stopping, because that's what happens once an agent leaves a demo and touches real systems. Most people only think about approval gates and plans when things go right. Watching your own team stumble and recover, even a simulated one, tends to change how you think about the ones you'd build for real.
Last updated 2 Oct 2026
More in Agent labs
- TRACE Agent EvaluationIs this task safe to hand to an autonomous agent at all?
- Build Your Own AI AgentHow does an AI agent actually work, and what makes one safe to launch?
