No single model is best at everything. A flock does not need one.
Levocat Flock is an open-source development environment I'm building, where models from different providers work as a team. Small squads take a task apart in parallel. When the problem is hard, they converge on it together.
Squads for breadth
A task splits into branches. An architect, a few builders and a reviewer each take one, in parallel, on the model that suits the job.
One flock for depth
When the problem is hard, the squads merge. Every model pulls in one direction, with one brief and one shared context.
A conductor holds the shape
It breaks the task into waves, gives every step a class of work, and hands that step to the model that has earned it.
The lead is earned, not assigned.
Nobody ranks models by hand. Flock benchmarks them on real tasks and lets a certified selector decide who leads each kind of work.
Measure
Run models and workflows on your own tasks, not on someone else's leaderboard.
Certify
A selection rule is fixed before the run and checked on held-out cases. It earns a certificate only if the evidence passes.
Route
A certified selector picks the model for every step, class by class. Where nothing is certified yet, your own order decides.
Pin
Want a specific model on a task? Pin it. A pin always wins over the selector.
Missing measurements are shown as unknown, never as a claim.
Why I'm building this.
Two ICLR 2026 papers, Trinity and Conductor, showed that a small coordinator can get more out of a team of models than any one of them gives alone. Flock is my way of bringing that idea into everyday development.
The smallest one in the room is in charge
Both papers in plain words, told the way a game developer would: turn-based tactics, raid leaders and fog of war.
Read on →The papers themselves: Conductor, Trinity. Levocat Flock is an independent project and is not affiliated with the papers' authors.
See every bird in the flock.
Under each message you see who worked on it: the conductor, its orchestrators, their subagents, each with a live status. Open any one and its transcript slides in beside your chat.
What's inside.
Many harnesses, one host
Agent harnesses from different providers run side by side. Choose one per chat, or let the conductor mix them inside a single task.
Work in waves
Steps that can run in parallel fly together. The flock regroups before the next wave, so results build on each other.
Agents name the work, not the model
Agents are plain files that point at a class of work. Whichever model has earned that class does the job.
Memory that lasts
Projects, history, editable memory and project knowledge carry across sessions, so you resume work with its context intact.
Benchmarks built in
Compare models and whole workflows on your own tasks by quality, execution time, and reported usage and cost.
Local by default
Chats, settings and project records stay on your machine. You choose which providers your agents work with.
Find your formation.
Levocat Flock is open source. Clone it, run it, make it yours.