auto-flowai
Decided to take another shot at multi-agent systems — built an engine for running a config-defined pipeline. Claude Code serves as the agent runtime. The engine itself is universal, but I chose to test-drive it on developing auto-flowai itself. The result is an engine plus a ready-made pipeline that emulates a typical project development workflow:
- PM writes the specification
- architect designs the solution
- tech lead approves the design, creates a branch and a PR
- developer and QA iterate on the code in a loop
- tech lead does the review
During execution the pipeline can improve itself through memory records and run history (initially I used a separate step with a meta-agent that edits the others’ instructions, but it made their instructions too narrow, which hurt codebase self-recovery).
Need to polish it up, turn it into a CLI utility, and try it on a couple of projects before Anthropic starts punishing such aggressive burning of subscription tokens.