They work alongside you in one chat: plan, research, write, code. They hand work to each other, run for real, and survive crashes. Built on Elixir and the BEAM.
No spam. Or follow the build on X @trile_412
↑ illustrative mock-up of a team chat, not a recording
Pick the agents you want and what each one may touch. Every agent gets its own capability list; joining a chat never grants tools.
team "weekly review"
agents: [planner, researcher,
writer, coder]
ask_me_before: :publicillustrative, not real syntaxYou and the agents share one thread. A planner splits the work; the others run in parallel as supervised BEAM processes, each in its own sandbox.
Results land back in the same chat. A failed turn shows up as failed; nobody writes text on an agent's behalf. You decide what ships.
Agents run long, they fail, and they talk to each other. That is what Erlang processes were made for: supervisors, no shared memory, and clustering out of the box.
Including the ones that went wrong. SQLite → Postgres one week in. WASI sandbox → bwrap. pgvector → plain-text memory.
Numbers straight from git, as of 2026-10-17. The repo is private for now; the write-ups aren't.
First commit, and a perf commit the same day: no O(n²), parallel tool calls, ETS hot paths.
One week in. Scheduler, triggers, retries and a secrets vault landed the same day.
The WASI sandbox lasted 10 days. Switched to bubblewrap with memory-aware admission.
Agents now write their own plain-text memory.
libcluster + Horde, plus a drain mode for deploys.
The UI said “failed” while the agent kept working. Runs now have owners and expiry times.
42 commits in one day, the busiest so far.
It lands on Oct 19. Until then, the daily log is on X @trile_412.
Leave your email and you'll hear when it opens.
You're on the list. I'll email you when it opens.
That didn't go through. Check the email and try again.
Or follow the daily log on X @trile_412