# How Cognition MultiDevin coordinates multiple agents

I wanted to understand how a well-known coding agent company coordinates many agents at once, using only what is public. I then asked the same about Claude Code to compare.

## What I found
- Cognition's release notes of September 2024 describe MultiDevin as one manager Devin and up to ten worker Devins. The manager gives each worker a task, then merges the changes from the workers that succeed into one branch or pull request. Maginative reported the same that month.
- In June 2025 Cognition's Walden Yan argued against building multi-agent systems. His post "Don't Build Multi-Agents" says that running several collaborating agents in 2025 gives fragile systems, because decisions get scattered and context is not shared well enough. The simplest design it names is a single-threaded linear agent. The study weighs coordination cost against the gain from specialization; the post makes no such comparison.
- The reason Cognition gives is implicit decisions. Its two principles: share context and full agent traces, not only single messages; and actions carry implicit decisions, so conflicting decisions give bad results. If two agents choose different patterns without seeing each other's choices, the parts do not fit together.
- Cognition's release notes say MultiDevin suits repeated, isolated tasks such as migrations, refactors and clean-ups. The study adds that tightly coupled features are a poor fit; Cognition does not say so. A later Cognition post (April 2026) says the original warning still holds for swarms of agents that write in parallel.
- Anthropic documents a different pattern for Claude Code. In its September 2025 post on the Agent SDK, subagents use their own isolated context windows and send back only relevant information. Current Claude Code documentation advises giving each parallel session its own checkout, such as a git worktree, so edits do not collide. The study describes the practice as a person choosing the boundaries and merging the work; Anthropic's pages do not say so.
- The study's sharpest contrast is who coordinates: in MultiDevin an agent decides how to split and merge, in the Claude Code practice a person does. The sources show the line has since blurred. In April 2026 Cognition describes a manager Devin that splits a larger task, starts child Devins and coordinates them through an internal MCP, and Claude Code's documentation now says Claude delegates to subagents on its own.
- What changed since January 2026: in March 2026 Cognition's post "Devin can now Manage Devins" describes a coordinator session, available to all users, that manages other Devins, each in its own isolated environment. It states no cap on their number, so the limit of ten from 2024 may no longer apply, and it does not use the name MultiDevin. Claude Code's documentation now describes a /batch command that splits a change into 5 to 30 units, shows the plan and, once a person approves it, gives each unit its own worktree.

## How it was made
I asked Perplexity Deep Research in January 2026 and it listed 38 sources, a mix of Cognition and Anthropic posts, press coverage and community write-ups. The AI gathered and drafted. I asked the question.

## Sources
- [Devin release notes 2024: Let Devin create Devins with MultiDevin](https://docs.devin.ai/release-notes/2024)
- [Cognition: Don't Build Multi-Agents](https://cognition.com/blog/dont-build-multi-agents)
- [Cognition: Multi-Agents: What's Actually Working](https://cognition.com/blog/multi-agents-working)
- [Cognition: Devin can now Manage Devins](https://cognition.com/blog/devin-can-now-manage-devins)
- [Maginative: Cognition Labs' Devin gets major upgrades](https://www.maginative.com/article/cognition-labs-devin-gets-major-upgrades/)
- [Jason Liu: Why Cognition does not use multi-agent systems](https://jxnl.co/writing/2025/09/11/why-cognition-does-not-use-multi-agent-systems/)
- [Anthropic: Building agents with the Claude Agent SDK](https://claude.com/resources/articles/building-agents-with-the-claude-agent-sdk)
- [Claude Code documentation: Best practices](https://code.claude.com/docs/en/best-practices)
