Agentic AI

The AI-Driven SDLC Harness

A software requirement enters as a work item. An orchestrator moves it through requirement analysis, design, implementation, review, human approval, testing and documentation. Each step leaves a persistent artifact. A human still owns the gates that matter.

The flow

1. Requirement agent turns the work item into a written requirement.
2. Design agent writes the design, with decision records when needed.
3. Developer agent makes the code change and its unit tests.
4. Review agent reviews the change. It cannot write source code.
5. Human approval. A model can recommend; it cannot approve.
6. Test agent. The test runner, not a model, is the source of truth.
7. Documentation agent records what changed.

The TaskFlow demo

The harness is the product. TaskFlow is the small task API it changes. The demo story, Add task priority support, has tasks gain Low, Medium or High priority, with Medium as the default; the API can update and filter by priority, and unit tests cover that behaviour.

Today the run is local and offline: requirement, design and developer agents each receive an explicit context bundle and return structured output through an offline stub, with no model, Jira or GitHub calls. The run stops at code review and persists its requirement, design and implementation artifacts, the context each agent saw, and the workflow state.

Design rules

• Agents communicate through artifacts and workflow state, not an open-ended conversation.
• The orchestrator chooses the next stage. A model can recommend; it cannot approve a merge or declare that tests passed.
• Context is assembled per agent, from paths that are named explicitly.
• Secrets stay in the environment; logs and error messages are scrubbed.

What comes next

Git branches and pull requests, then reading the story from Jira, then review, the human gate UI and the test agent. A public demo will show a stored run rather than handing visitors live credentials.

TODO: content from Alex — screenshots or a stored run of the TaskFlow demo, and a link to the repository if it becomes public.