Multi-Agent Wiki
Environment

Role-playing / SOP / Virtual Company

Agents play product, architect, dev, QA roles bound by a documented SOP.

Live visualizationAnimated topology — press Space to play / pause
·Diagram
Live
1 / 8

—

Definition

Assign agents professional roles — product, architect, dev, QA — and constrain their collaboration with a Standard Operating Procedure.

Category: Execution environment

When to use

Virtual software companies, product design, teaching simulations, organizational process automation.

When not to use

When roles are decorative — no tools, no acceptance criteria.

How to implement

  1. Each role declares responsibility boundary, allowed tools, and input/output formats.
  2. The SOP is encoded as a workflow, not just buried in a prompt.
  3. Role handoffs produce artifacts, not just chat logs.
  4. For software-development tasks, plug in real test execution and code running.

Minimal pseudocode

TypeScript
const sop = [
  { role: "PM", output: "requirements.md" },
  { role: "Architect", output: "design.md" },
  { role: "Engineer", output: "patch.diff" },
  { role: "Tester", output: "test-report.md" },
];

for (const step of sop) await role(step.role).run(step);

Recommended trace events

  • role.task.started
  • role.artifact.created
  • sop.step.completed
  • sop.violation.detected

Common failure modes

  • Agents play the role well but produce no real output.
  • The SOP balloons in length and cost.
  • Role outputs don't follow a consistent artifact format.

Implementation checklist

  • Input/output schemas defined.
  • Each agent's permission boundary defined.
  • Every agent call carries a run id / trace id.
  • Failure, timeout, cancel, and retry strategies defined.
  • Context passed is the minimum required, not the full history.
  • High-risk actions are gated by approval or a verifier.

References