The problem
What does the development process look like when a product change begins as a conversation and an agent carries out the implementation?
This project explores that workflow, including the infrastructure, permissions, validation, and human decisions needed around the coding agent.
Architecture
The high-level flow moves from a feature request through requirements and a specification, into a development workflow, and finally to a pull request for human review.
- 01User
- 02Feature request
- 03Conversation / requirements
- 04Specification
- 05Agentic development infrastructure
- 06Coding agent
- 07Branch
- 08Implementation
- 09Automated tests / validation
- 10Pull request
- 11Human review
- 12Merge
The central architectural principle is that the human currently enters the loop primarily at the pull-request boundary.
Technical decisions
The pull request is the current boundary between autonomous implementation and human approval. It gives the human reviewer a concrete change to assess, alongside the available tests and validation.
The right constraints around this workflow are an active area of exploration:
- Where should human approval occur?
- How should agent permissions be constrained?
- How should coding environments be isolated?
- How should agent activity be logged?
- How should generated code be evaluated?
Specific infrastructure and technology choices will be documented here as the case study develops.
Questions driving the work
- What work can safely be delegated entirely to agents?
- What work requires hybrid human and agent development?
- What work should remain human-led?
- How does business exposure affect acceptable autonomy?
- How does security risk affect acceptable autonomy?
- What happens when agents produce code faster than humans can review it?
Lessons learned
This project is active and experimental. Findings, implementation details, and more detailed architecture diagrams will be added as they are ready to share.