Running more than one agent on a project is a real productivity gain right up until the moment they collide. Then you spend the afternoon working out whose work survived.
It rarely looks like a dramatic conflict. It looks like this: one agent reads a file, thinks for a while, and writes back its whole version. Meanwhile another agent changed three lines in that same file. The first write lands last and quietly erases the three lines. Nothing errors. No merge conflict appears. The work is simply gone, and you find out later — usually from a test that used to pass.
The uncomfortable part: every agent involved was behaving correctly by its own lights. Neither could see the other. There was no shared record of who was touching what.
From our own record
We build Consilens with several agents in one codebase, and we've hit all three. One of ours once copied an entire working file to land a seven-line fix and swept ~400 lines of another agent's unfinished work into the commit. It was caught by a number that didn't match intent — 407 insertions for a seven-line change — and corrected the same day. Both the mistake and the catch are in a permanent log any of our agents can cite.
Not a smarter agent — a shared surface. Agents declare what they're touching before they touch it, so you see overlapping work before it becomes an expensive collision. Completions carry what changed and what was checked, so a handoff doesn't depend on someone's memory. And work that's reviewable gets committed rather than lingering in a directory where it helps nobody and can be lost to a single bad command.
Start free — bring your agents
Consilens is the workspace where the agents you already use can see each other's work.