ForgeFlow
AI agents that read your GitHub issues and ship pull requests autonomously.
Overview
A multi-agent orchestration platform that takes a GitHub issue, breaks it into tasks, writes code, runs QA, and opens a pull request — with human oversight at every stage.
The Problem
What wasn't working
Senior engineers burn hours on routine issues — bug fixes, boilerplate features, dependency updates — work that's well below their pay grade. Every hour spent on grunt work is an hour not spent on architecture, product decisions, or shipping features that move the needle.
The Solution
How I solved it
I built a pipeline of specialized AI agents — Director, Coder, QA, and Farmhand — that collaborate to resolve issues end-to-end. A Supervisor cycle (Observe → Diagnose → Decide → Record → Enforce) keeps agents on track with configurable budgets and escalation policies.
Process
Step by step
Designed the multi-agent architecture with clear role boundaries
Built each agent with focused capabilities and configurable budgets
Implemented the 5-stage Supervisor cycle for quality enforcement
Created a real-time HTMX dashboard for monitoring agent activity
Added human-in-the-loop escalation for edge cases
Tested across diverse repository types and issue complexity levels
Challenges
What made it hard
Preventing agents from going off-track required building an Opposed Director agent that challenges decisions before execution.
Handling codebases of varying sizes and languages meant building flexible context management that adapts to each repository.
Results
The impact
Autonomous issue resolution
Routine issues go from open to merged PR without human intervention.
Multi-agent pipeline
5 specialized agents with configurable roles and budgets.
Human oversight built-in
Escalation policies ensure a human reviews what matters.
45 tests passing
Comprehensive test coverage across 8 test modules.
Gallery
Visual walkthrough
Agent Pipeline
Live Dashboard
Supervisor Cycle