Cognition demos "Agent Fan Out" to run 10 Devins in parallel

Jared Zoneraich, a builder-in-residence at Cognition, presents "Agent Fan Out" as the most impactful pattern he has learned from Cognition's internal engineering team. The core idea is that instead of a single agent grinding through a large task sequentially, a lead Devin agent breaks the work into discrete chunks, spins up ten or more sub-Devin sessions — each running in its own VM — to execute those chunks in parallel, and then the lead agent merges all results. Zoneraich notes that Cognition's own research team uses a version of this at scale, spinning up a hundred Devins to read different parts of training logs simultaneously, since context window constraints make it more effective to have each agent focus on one specific thing.

The video demonstrates the pattern live by building a Sicily travel app.

The video demonstrates the pattern live by building a Sicily travel app. Devin is prompted to research Palermo across several categories (restaurants, language, maps, points of interest, budget, and others it chooses itself), then spin up a separate sub-Devin for each category to do deep research and push results to a shared Git repo. Zoneraich emphasizes letting Devin write the prompts for its own sub-agents, describing the generated prompts as better than what he would write himself. Additional techniques shown include instructing agents to work in parallel whenever tasks are separable, running iterative design loops inside sub-sessions, and spinning up three sub-Devins to build the same feature three different ways in separate branches — treating engineering effort as cheap enough to build multiple versions and keep the best.

Key facts

1. 01The pattern is called 'Agent Fan Out': a lead Devin spawns 10 or more sub-Devin sessions that work in parallel, then reports back to the lead for merging.
2. 02Each sub-Devin runs in its own VM but sub-agents can communicate with each other.
3. 03Cognition's research team uses the pattern at scale, spinning up a hundred Devins to read different sections of training logs simultaneously.
4. 04The live demo builds a Sicily travel app, with Devin spinning up seven sub-agents to research separate categories (food, language, maps, points of interest, budget, and others).
5. 05Zoneraich recommends letting Devin write the prompts for its own sub-agents, saying the generated prompts are better than human-written ones.
6. 06A key technique shown is spinning up three sub-Devins to build the same feature three different ways in separate branches, then keeping the best result.
7. 07Presenter is Jared Zoneraich, identified as a builder-in-residence at Cognition.

Topics