Blinkin Legacy Docs Docs
Manuals · manuals/09-agent-team-with-apps.md Docs Home

Manual 09: Create A Team Of Agents With An App

What This Guide Is For

Use this guide when one Agent is not enough and you want an App to coordinate several Agents.

The App becomes the workflow. Each Agent has a role. The App passes work from one Agent to the next.

What You Are Building

Example:

Intake Form
-> Planner Agent
-> Research Agent
-> Reviewer Agent
-> Writer Agent
-> Human Approval
-> Publish / Export / Send

This is an agent team. It is useful when a task needs multiple skills, review, or staged output.

When To Use An Agent Team

Use an Agent team when:

  • one Agent tries to do too many different jobs;
  • you need research plus writing plus review;
  • one Agent needs limited Tools and another needs different Tools;
  • the result should pass through approval before publishing;
  • you want a repeatable workflow, not just an open chat.

Do not use an Agent team when a single clear Agent can complete the task.

Before You Start

You need:

  • permission to create Apps;
  • permission to create Agents;
  • a clear workflow goal;
  • at least two Agents with different roles;
  • optional: Spaces, Tools, integrations, or webhooks.

Step 1: Define The Team Roles

Write down the Agents you need before creating the App.

Example roles:

  • Planner Agent: turns user input into a task plan.
  • Research Agent: searches sources and gathers evidence.
  • Analyst Agent: compares options or extracts structured facts.
  • Reviewer Agent: checks quality, risks, and missing information.
  • Writer Agent: turns approved material into final output.
  • Publisher Agent: formats or sends the result.

Expected result: each Agent has one clear responsibility.

Step 2: Create Or Select The Agents

  1. Go to Agents.
  2. Create each Agent or choose existing Agents.
  3. Give each Agent a specific instruction.
  4. Enable only the Tools that Agent needs.
  5. Connect only the Spaces that Agent needs.
  6. Test each Agent alone.

Expected result: each Agent works by itself before being added to the App.

Step 3: Create The Coordinating App

  1. Go to Apps or Agentic Apps.
  2. Select Create.
  3. Choose Agent Team, Workflow App, or Blank App.
  4. Add a name and short description.

Example:

Name: Market Research Team
Description: Plans, researches, reviews, and drafts a market research brief.

Expected result: an App draft is ready to coordinate the Agents.

Step 4: Add The Intake Step

  1. Add a first Step with a form.
  2. Collect the user request.
  3. Add fields for constraints such as region, audience, deadline, format, and files.
  4. Save.

Expected result: the App captures the information the Agent team needs.

Step 5: Add The Planner Agent Step

  1. Add an Agent Step.
  2. Select the Planner Agent.
  3. Map the intake fields into the Planner Agent.
  4. Define the Planner output as a structured task plan.

Example output:

goal
questions_to_answer
sources_to_check
risks
next_agent_instructions

Expected result: the App creates a plan before work begins.

Step 6: Add Worker Agent Steps

Add one Step per specialist Agent.

Example:

Research Agent receives: task plan + source requirements
Analyst Agent receives: research notes + comparison criteria
Writer Agent receives: approved findings + output format

For each Agent Step:

  1. Select the Agent.
  2. Map previous Step outputs into that Agent's input.
  3. Define what the Agent must return.
  4. Save.

Expected result: each Agent receives only the context needed for its role.

Step 7: Add A Reviewer Agent

  1. Add a Reviewer Agent Step after the worker Steps.
  2. Give it the draft output and original user request.
  3. Ask it to check:
    • factual support;
    • missing requirements;
    • risky claims;
    • formatting issues;
    • whether human review is needed.
  4. Save.

Expected result: the App can catch quality problems before final output.

Step 8: Add Human Approval

  1. Add a review Step for the user or admin.
  2. Show the Agent team result.
  3. Add actions:
    • approve;
    • request changes;
    • retry Agent;
    • send to another Agent;
    • publish/export.

Expected result: important outputs are not published silently.

Step 9: Add Final Action

Choose what happens after approval.

Options:

  • create a document;
  • publish an App result;
  • send an email;
  • call a webhook;
  • update an integration;
  • export a PDF or table;
  • save to a Space.

Expected result: the Agent team's work produces a durable output.

Step 10: Test The Full Team

  1. Run the App with a simple input.
  2. Check each Agent Step.
  3. Confirm each handoff has the right data.
  4. Check that failed Agent Steps can be retried.
  5. Check human approval.
  6. Check the final action.

Expected result: the Agent team completes a full workflow without manual copy/paste between Agents.

Common Team Patterns

Research Team

Planner -> Researcher -> Source Reviewer -> Writer -> Approval

Document Extraction Team

Intake -> OCR Agent -> Data Normalizer -> Reviewer -> Export

Sales Enablement Team

Company Intake -> Research Agent -> Messaging Agent -> Reviewer -> CRM/Webhook

Support Triage Team

Issue Intake -> Classifier Agent -> Knowledge Agent -> Response Agent -> Human Approval

Compliance Review Team

Upload -> Extractor Agent -> Policy Agent -> Risk Reviewer -> Report

Good Agent Team Design

  • Give every Agent one job.
  • Keep Tool access narrow.
  • Use structured outputs between Agents.
  • Add a Reviewer Agent for important results.
  • Add human approval before public publishing or external sending.
  • Store final outputs in a durable place.

Troubleshooting

  • If the team gives messy results, make each Agent's output more structured.
  • If one Agent repeats another Agent's work, narrow its instructions.
  • If the workflow is slow or expensive, remove unnecessary Agent Steps.
  • If a later Agent lacks context, check output mapping from the previous Step.
  • If private data leaks into the wrong Agent, check Space and Tool access.
  • If final output is weak, add a Reviewer Agent or human approval Step.

Product Decisions Needed

  • Exact label for this pattern: Agent Team, Crew, Workflow App, or Agentic App.
  • Whether Agent teams are created from templates.
  • Whether every Agent Step must have a typed input/output contract.
  • Whether public Agent teams require review before publication.
  • How Agent team runs appear in logs and billing.