No. 06
Author · 2026 - Present
claude-agent-workflow
Open-source reference implementation of an agent-orchestrated development workflow

Context
claude-agent-workflow is a public, MIT-licensed reference implementation of an agent-orchestrated development workflow. It extracts the pattern of a workflow I rolled out at Solvexus; the code is written from scratch, contains no proprietary code, and is meant to be read and reused.
Role
Author and sole contributor.
Tasks
- Show the workflow as working code instead of a description.
- Keep the agents' roles separate and their hand-offs explicit.
- Block unverified output with gates.
What I did
- Split the work into investigator, per-stack implementer and reviewer roles with explicit hand-off contracts.
- Added quality-gate hooks that fail closed.
- Wrote a governed-TDD skill for the red, green, refactor cycle.
- Wrote a tracker MCP server in Python.
Stack
- Claude Code
- MCP
- Python
- Shell
Result
The repository is public on GitHub under the MIT license. It comes from the workflow I designed and rolled out at Solvexus, with 12 per-stack implementer, reviewer and investigator agents, adopted as the standard way of working by the entire engineering team; there I also built MCP servers for Azure DevOps, Azure Application Insights and Jam.