Open-source Slack for your AI agent swarm
Python
# One agent owns the code and alone pushes to the shared branch. The others send it tests and patches
# through a second repository, /incoming/repo.git, which every agent may push to.
name = "owner8"
tasks = ["sitkevij__hex.61ae69b"] # ProgramBench task ids. README.md lists the ten in the blog post.
hours = 6 # agent time per task
snapshot_minutes = 30 # save the submission this often, to score it over time. 0 for never.
parallel = 1 # tasks run at once
loop = true # each time an agent ends its turn early, send it back to work
[agent]
# The agent's command-line tool, a Linux x86_64 build. It is mounted read-only into each container.
kind = "claude" # claude, codex or command
binary = "~/.local/bin/claude"
model = "claude-opus-5-5"
effort = "xhigh"
# For Codex instead:
# kind = "codex"
# binary = "~/agents/codex"
# model = "gpt-6-luna"
# effort = "xhigh"
[team]
members = 8
[[team.roles]]
name = "owner"
count = 1
prompt = "../prompts/owner.md"
loop = "../prompts/owner-loop.md"
[[team.roles]]
name = "helpers"
push = false # the shared repository is read-only for these agents
prompt = "../prompts/helper.md"
loop = "../prompts/helper-loop.md"