Inspiration
GPT-5.6 surprised us with how capable it is at spawning and coordinating multiple AI agents. However, agents still need a reliable orchestration layer when they share a codebase.
Git worktrees are not always a good solution. They can lack required environment variables, create development-server port conflicts, consume significant resources, duplicate dependencies, and require cleanup. We wanted a tool that would let us work on many different issues concurrently on the same branch and in the same working tree.
What it does
Bandmaster is a local CLI and Codex skill that coordinates multiple agents in one Git working tree. It assigns tasks, gives workers exclusive ownership of files, preserves progress and handoffs, detects unsafe changes, validates completed batches, and creates one attributable commit per task.
How we built it
We built Bandmaster in Go as a local-first CLI. SQLite stores durable sessions, tasks, leases, file claims, validation results, and audit history under Git metadata. A filesystem monitor and authoritative repository scans detect unexpected changes, while a central state machine manages the workflow from planning and assignment through validation and commit.
Bandmaster also generates a project-level Codex skill that teaches the parent agent and its workers how to use the orchestration protocol. A Bubble Tea TUI provides a live view of the current session, tasks, ownership, and integrity status.
Challenges we ran into
The greatest challenge was safely coordinating agents that run as the same operating-system user and edit the same working tree. Filesystem permissions cannot isolate them, and a single unclaimed edit can compromise the entire session. We addressed this with atomic file claims, worker leases, continuous monitoring, and authoritative integrity checks at every important transition.
Recovery was equally difficult. Workers can disappear, validation commands can mutate files, Git hooks can fail, and commits can stop halfway through.
## Accomplishments that we're proud of
Our proudest accomplishment is that we used Bandmaster to improve Bandmaster. It is the first project where we created a tool capable of coordinating agents working on its own codebase. Seeing multiple agents safely build and improve the system responsible for orchestrating them was an amazing moment.
We also built a complete coordination lifecycle: durable planning, atomic file ownership, leases, structured handoffs, batch barriers, validation, recovery, and deterministic commits. An extensive integration suite tests these guarantees through the public CLI against real Git repositories.
What we learned
We learned that GPT-5.6 is remarkably capable of spawning, coordinating, and managing AI agents. The limiting factor is no longer simply the model’s intelligence, it is the orchestration infrastructure around it.
Codex needs a better way to track ownership, communicate progress, preserve context, coordinate plans, and recover from interruptions. Parallel agents need more than delegation: they need durable shared state, explicit responsibilities, clear barriers, and safe recovery paths.
What's next for Bandmaster
Next, we want to improve the TUI and provide clearer, more actionable diagnostics when something goes wrong.
We also want to give AI agents richer ways to interact through Bandmaster. This includes better agent-to-agent communication, collaborative planning, progress reporting, dependency management, and recovery tools. Our goal is for Bandmaster to become a shared coordination layer that helps groups of agents reason and work together, not merely avoid editing the same file.
Log in or sign up for Devpost to join the conversation.