Agentic Workflow is a portable operating model for AI-assisted software delivery. It gives Copilot, MCP-enabled IDEs, and command-line coding agents a shared contract for planning, implementation, debugging, review, and durable memory.
The first version of this repo was a set of specialist prompts. This version turns it into a usable kit:
- A versioned workflow manifest:
workflow/default.workflow.json - Specialist prompt definitions:
specialists/*.xml - Copilot and agent instruction templates:
templates/ - A dependency-free CLI for validation and installation:
bin/agentic-workflow.mjs - CI validation:
.github/workflows/validate.yml - Adoption and operating docs:
docs/
Agentic development does not need more vague roleplay. It needs handoffs that survive context resets, reviewers that see evidence, and memory that stores decisions instead of transcripts.
Agentic Workflow is built around five roles:
| Role | Job | Output |
|---|---|---|
| Architect | Shape ambiguous work into a scoped plan | Acceptance criteria, plan, risks |
| Implementer | Make the smallest coherent change | Diff, docs, verification notes |
| Debugger | Prove behavior and isolate failures | Reproduction, root cause, check output |
| Reviewer | Find correctness and maintainability risks | Ordered findings and residual risk |
| Historian | Preserve durable learning | Decisions, outcomes, follow-ups |
Each role has a handoff target and quality gates in the manifest. The workflow is complete only when the next role has enough evidence to continue without replaying the full chat.
Validate this repository:
npm testInstall the workflow into another repository:
node ./bin/agentic-workflow.mjs init /path/to/your/repoValidate an installed repository:
node ./bin/agentic-workflow.mjs validate /path/to/your/repoRender a role prompt from the manifest:
node ./bin/agentic-workflow.mjs render-prompt architectinit writes the portable workflow surface into a target repository:
.agentic/workflow.jsonAGENTS.md.github/copilot-instructions.md.github/instructions/tests.instructions.md.github/prompts/architect.prompt.md.github/prompts/implementer.prompt.md.github/prompts/debugger.prompt.md.github/prompts/reviewer.prompt.md.github/prompts/historian.prompt.md.vscode/mcp.json
Use --force to overwrite existing files intentionally.
.
├── bin/
│ └── agentic-workflow.mjs
├── docs/
│ ├── adoption-guide.md
│ └── operating-model.md
├── examples/
│ └── .agentic/session.md
├── specialists/
│ ├── architect.xml
│ ├── implementer.xml
│ ├── debugger.xml
│ ├── reviewer.xml
│ └── historian.xml
├── templates/
│ ├── AGENTS.md
│ ├── .github/
│ └── .vscode/
└── workflow/
├── agentic-workflow.schema.json
└── default.workflow.json
As of 2026-04-30T20:41:26-05:00, current official documentation supports the direction of this repo:
- GitHub Copilot supports repository instructions, path-specific instructions, prompt files, and agent instruction files such as
AGENTS.md. - VS Code recommends file-based custom instructions and prompt files over older settings-based generation instructions.
- MCP defines prompts as user-controlled reusable templates and tools as model-controlled external actions with a human approval surface.
Agentic Workflow uses those native surfaces instead of requiring a proprietary runtime.
The repo keeps .vscode/mcp.json as a starter configuration. The install template is intentionally minimal:
- Context7 for current documentation lookup.
- Playwright MCP for browser/UI verification.
Add Octocode, PAMPA, Clear Thought, or other MCP servers when your team has a concrete evidence need for them. Keep tokens and credentials outside the repository.
This project intentionally has no runtime dependencies.
npm run validate
npm run smoke
npm testApache-2.0