A configurable generator and template set for scaffolding specialized AI agent roles around a visible GitHub Issues work queue and outcome gates.

The workflow before the build

Complex work can arrive as an unstructured request with unclear ownership, hidden handoffs, and no durable record of what each specialist contributed.

What the public project delivers

The public starter kit packages a Python generator, configurable persona and workflow templates, GitHub Issues and Actions patterns, and Azure Functions and Azure OpenAI integration scaffolding.

  • Generates configurable specialist persona and workflow files.
  • Uses GitHub Issues and Actions templates for a durable, visible work queue.
  • Provides scaffolding for Azure Functions and Azure OpenAI integration.
  • Includes outcome framing, validation, and human intervention patterns.

The value case, stated honestly

The pattern is designed to make multi-agent work easier to assign, inspect, and govern while keeping the work item and its history visible to people.

Evidence label: Designed value: clearer ownership, inspectable handoffs, and human-visible outcome gates for multi-agent work.

Architecture in one view

A GitHub issue moves through workflow templates and generated persona scaffolding to a human outcome gate.

Agent Team Starter Kit architecture A GitHub issue moves through workflow templates and generated persona scaffolding to a human outcome gate. 01 GitHub issue 02 Workflow templates 03 Generated personas 04 Human gate
01 / Issue

A GitHub issue holds the request, context, and visible work state.

02 / Workflow templates

GitHub Actions and issue-handling templates describe routing and intervention points.

03 / Generated personas

Configurable specialist files scaffold bounded roles that deployers must complete and wire.

04 / Human gate

A person can inspect the work record and decide whether the outcome is acceptable.

Human control and guardrails

  • GitHub keeps the request, discussion, and resulting work visible rather than hiding them in a transient chat.
  • Specialist roles and outcome gates can be configured for the workflow instead of granting one agent an undefined remit.
  • People remain responsible for reviewing consequential outputs and deployment choices.

Where the pattern stops

  • This is scaffolding rather than turnkey autonomy: generated personas include TODOs, the runner is a stub, and deployers must complete the domain logic and wiring.
  • Setup requires Python, authenticated GitHub CLI access, Azure Storage, Azure OpenAI, and secured repository permissions.
  • Evaluation, monitoring, cost controls, and domain-specific safety checks still need to be designed for each use.

Explore the public proof