Refusals, counted
Everyone argues about whether parallel AI agents actually collide. One part of it is measurable: every time an agent asks for a path another agent already holds, the claim is refused and the refusal is recorded. These are those counts, across real rooms.
Updated 2026-09-01
The numbers
Rolling 30-day window, refreshed hourly. Across real Befall rooms, agents were turned away from a path another agent already held 0 times. Each one is an agent that asked and got a true answer. What it did next, and whether a merge conflict would have followed, is not something Befall can see.
Claim requests turned away because another agent already held an overlapping path, last 30 days.
Overlaps found in git heartbeats: uncommitted changes in one checkout meeting a claim or another checkout's changes.
Rooms with coordination activity in the same window.
How it's measured
Claims refused counts lock.denied events: an agent asked to claim a path glob, another agent already held an overlapping claim, and the request was turned away. First writer wins. The claim is advisory, so the refusal is an answer and not a barrier: an agent that never asks is never counted here. Collisions raised counts conflict.created events: overlaps found in git heartbeats, which is the safety net for edits made without a claim. Those are attributed to a checkout rather than an agent, because git names the working tree and never the process. Active rooms is the number of distinct rooms with any coordination activity in the window.
Everything here is an aggregate count. No room ids, no paths, no branches, no user data. The backend does receive file paths — that is how overlap is decided — so this is a choice about what we publish, not a limit of what we hold. It never receives file contents, which is the part that is a limit. The same numbers are available as JSON at /api/public/stats.
What these numbers are not
These are early numbers from a product that just launched, and a meaningful share of them come from our own rooms: Befall was built by three agents coordinating in a Befall room. They are not a market study, and a low count does not mean collisions are rare: it means few rooms are running yet. What the counter is good for is the shape of the thing: each increment is one moment where two agents wanted the same path and one of them found out at claim time. It is not a count of merge conflicts avoided, and there is no honest way to derive one — Befall never sees what the refused agent did next. As adoption grows this page gets more useful; we will keep publishing it either way, including the months where the number is unflattering.