Guide
Devin (Cognition) guide: a practical look at the autonomous coding agent
By Atlas AI Resources · Published November 18, 2026
Devin is Cognition's autonomous coding agent. You give it a repository and a task, and it works on its own cloud machine with a shell, an editor, and a browser, then hands back a pull request. This page summarizes what the Devin docs and the pricing page say, then adds our own advice for running it next to other agents.
1. What it is
- An agent with its own machine. The docs describe Devin as an autonomous AI software engineer that can write, run, and test code. Each session gets a workspace with a Shell, an IDE, and a Browser, and you can watch or take over any of them.
- Sized for a few hours of work. Cognition's rule of thumb: if a task would take you three hours or less, Devin can most likely do it. Bigger projects should be split into focused sessions.
- Many ways in. The web app at app.devin.ai, Slack, Linear, a
/devincomment on a GitHub pull request, the Devin API, and the Devin CLI, which can hand a longer task off to a cloud session with/handoff. - More than coding sessions. The same account includes Ask Devin (questions and planning), Devin Review (pull request review), DeepWiki, and Automations. This page sticks to the coding agent.
2. Per the docs: how a session works
- Ask, then Agent. A session starts in one of two modes. Ask explores the codebase and plans without changing code; Agent writes code, runs commands, browses, and opens pull requests. Unless the task is already fully scoped, the docs recommend starting in Ask, building a prompt from it, and selecting Send to Devin.
- Pick a repository and an agent. The repository selector sets the codebase and starting branch. The agent picker offers modes such as Lite, Normal (the default), Ultra, and Fusion; higher capability costs more. Admins control which ones you see.
- Give it context with @ mentions. You can point at repositories, files, skills, secrets, and earlier sessions.
- Watch and steer. The Progress tab, Shell, IDE, and Browser show what Devin is doing. Side chats let you ask questions about the session without interrupting it; they're read-only. To take over, you stop the session and use the IDE yourself.
- Pull requests and comments. Devin responds to comments on pull requests its sessions are working on, as long as the session isn't archived. On any open pull request in a connected repository, a comment starting with
/devinstarts a session, and/devin reviewstarts a Devin Review. - Sleep. An idle session goes to sleep after 5 minutes of inactivity by default and doesn't use quota while asleep. Send a message to wake it.
- After the session. Session Insights gives a timeline, feedback, and an improved prompt you can reuse for similar tasks.
3. Per the docs: setup steps
- Sign up at app.devin.ai and pick a plan (see section 5).
- Connect your git host. In Settings, then Connections, connect GitHub, GitLab, Bitbucket, or Azure DevOps. For GitHub you can grant All repositories or Select repositories; you need to be an admin of the GitHub organization.
- Set up the environment. Every session boots from a snapshot of a Linux virtual machine with your repositories cloned and dependencies installed, and session changes don't persist back to it. The simplest path is to start a session and ask "Set up your environment for this repo." Devin proposes a blueprint, you approve it, and a build creates the snapshot. The docs call this the single highest-leverage thing you can do for Devin's results.
- Add secrets the right way. Store credentials as encrypted Secrets and reference them as variables; Devin can't discover your credentials on its own. Organization-wide secrets are usable in every future session, and by default any member can manage them.
- Write down how your repository works. Devin reads an
AGENTS.mdfile and automatically includes the first 16 KiB of it, so keep the important rules near the top. For reusable procedures, use Skills (SKILL.mdfiles). Playbooks and Knowledge are being moved to Skills, so start new instructions there. - Connect where work comes from. In Slack, tag @Devin and it replies in the thread. In Linear, assign a ticket to Devin or add a skill label. Automations can start sessions from Slack, GitHub, Linear, schedules, and webhooks.
- Optional: the API. The current v3 API uses service users with role-based access; the v1 and v2 APIs still work but get no new features. See the API overview.
4. What it's good and bad at
Per the docs, Devin does best on:
- Linear or Jira tickets, bug reproduction and fixes, unit tests, and documentation upkeep.
- Migrations and refactors, such as JavaScript to TypeScript or framework upgrades.
- Parallel batches of independent tasks you return to as draft pull requests.
- Tasks it can check itself: tests, lint, a compile step, or CI going green.
Where it struggles, also per the docs: extremely difficult tasks, vague prompts without completion criteria, and repositories whose environment isn't set up. If Devin keeps hitting session usage limits, the docs say the task is probably too complex and should be split. The docs also note Devin may not behave exactly as described and that the docs may be out of date, so test on your own repository before you rely on it.
5. Cost basics (from the pricing page)
Prices as listed on devin.ai/pricing on November 18, 2026. Check that page before you buy; Cognition changed its plan lineup in April 2026.
- Free: $0, with a light usage quota and limited model availability.
- Pro: $20 a month, one user, with access to cloud agents (Devin Cloud) and the option to buy extra usage at API pricing.
- Max: $200 a month, one user, with much higher quotas.
- Teams: $80 a month, billed as prepaid credits shared by the whole team, up to 100 users, with SSO and cost controls.
- Enterprise: custom pricing, billed in Agent Compute Units (ACUs) at the rate in the order form.
How usage works: Pro has a daily and weekly allowance and Max a weekly one. Past that, prepaid on-demand credits keep sessions running; if they run out too, sessions pause. Usage grows with the number and complexity of Devin's actions, session length, and back-and-forth. Devin Review and Automations also draw on usage. Short, clearly scoped sessions cost less.
6. Devin vs Jules, Codex, Copilot, and Cursor
All five follow the same basic pattern: an isolated machine, a prepared environment, and a branch or pull request that waits for a person. The table only lists what our guides and the vendors' docs support.
| Agent | Where you start it | What stands out |
|---|---|---|
| Devin | Web app, Slack, Linear, /devin PR comments, API, CLI | Ask mode to plan first; a snapshot-based environment you can set up by asking |
| Google Jules | Web app, a "jules" GitHub issue label, CLI, API | Writes a plan before it works; scheduled and suggested tasks |
| OpenAI Codex Cloud | Web, ChatGPT desktop app, mobile | Reusable cloud environments that Codex prepares and tests |
| GitHub Copilot cloud agent | github.com, GitHub Mobile, an IDE, Copilot CLI | Built into GitHub; code review and fixes for failed workflows |
| Cursor Cloud Agents | Cursor desktop, web, Slack, GitHub, Linear, API | Own VM per agent, results come back as pull requests |
Whichever you use, put the real gate in your repository settings, not the prompt. Devin's GitHub docs recommend branch protection on your main branch so required checks pass before Devin can merge changes.
7. Our advice: running it in a small agent fleet
- One job per agent. Give Devin one kind of work, such as bug tickets from one Linear label or test coverage on one repository, and leave other jobs to other agents. It keeps costs readable and makes failures easy to trace.
- Write tasks like tickets. A clear start and end, the files or patterns to follow, and a success check Devin can run. "Make this better" isn't a task.
- Plan before you spend. For anything bigger than a small fix, use Ask mode first and send a scoped prompt to Agent mode.
- Review every pull request. Keep a person as the merge step, enforced by branch protection and required reviews. Green CI isn't the same as correct; read the diff.
- Keep secrets narrow. Use a dedicated account for Devin, as its docs suggest, and add only the secrets the job needs, since organization secrets reach every session.
- Judge by what you keep. Count merged, kept pull requests and what each one cost, not sessions started. Add a second task type only once the first is boring.
Related: Google Jules guide · OpenAI Codex Cloud guide · GitHub Copilot cloud agent guide · Cursor Cloud Agents guide · Claude Code guide · Small AI fleets: handoffs, group chats, and the shared computer · Keeping AI agents safe from prompt injection · The Usage Limit Is Why My Grok Bots Have Jobs