Skip to content

operating-model.md

Latest commit

 

History

History
54 lines (36 loc) · 2.7 KB

operating-model.md

File metadata and controls

54 lines (36 loc) · 2.7 KB
Raw

Operating Model

Agentic Workflow is a coordination layer for AI coding tools. It does not try to be another agent runtime. Its job is to make agent work observable, repeatable, and portable across Copilot, MCP-enabled IDEs, and command-line agents.

Why This Exists

Most agentic coding failures are not caused by a missing prompt trick. They come from weak handoffs:

  • The planner does not define acceptance criteria.
  • The implementer changes scope without saying so.
  • The debugger reports confidence without reproduction evidence.
  • The reviewer comments on style while missing behavior risk.
  • The final summary loses the decision trail.

This project gives those handoffs an explicit shape.

Core Artifacts

  • workflow/default.workflow.json is the source of truth for roles, phases, and quality gates.
  • specialists/*.xml keeps compatibility with prompt-copy workflows.
  • templates/AGENTS.md gives coding agents a durable repo-local briefing.
  • templates/.github/copilot-instructions.md and prompt files integrate with GitHub Copilot customization.
  • bin/agentic-workflow.mjs validates the kit and installs it into another repository.

Phase Contracts

Each phase has an output contract. A phase is not complete because an agent says it is complete; it is complete when the next phase has enough evidence to continue.

Phase Owner Main output
Shape Architect Acceptance criteria and implementation plan
Build Implementer Focused diff and verification notes
Prove Debugger Reproduction, root cause, and check output
Review Reviewer Ordered findings and residual risks
Remember Historian Durable decisions and follow-ups

Quality Gates

The default workflow uses four gates:

  • Scope: a bounded goal and explicit non-goals.
  • Context: inspected code and current docs when needed.
  • Verification: tests, build, typecheck, or deterministic manual proof.
  • Handoff: compact state for the next agent.

Teams can extend these gates in .agentic/workflow.json.

Relationship to MCP

MCP tools and prompts are useful, but they are not a workflow by themselves. The official MCP prompt model exposes reusable prompt templates that users can discover and invoke, while MCP tools let models call external systems with a human approval surface. Agentic Workflow sits above that layer: it decides what evidence a role needs before using tools and what artifact must be handed off afterward.

Relationship to Copilot Instructions

GitHub and VS Code support repository instruction files, path-specific instructions, prompt files, and agent instruction files. Agentic Workflow ships templates for these native surfaces instead of inventing a proprietary prompt loader.