Autonomous coding agents are the most ambitious category in AI tooling: give the system a ticket, walk away, come back to a pull request. No autocomplete, no chat - the agent plans, codes, tests, and iterates on its own, often in a cloud sandbox with its own shell, editor, and browser.
Three tools define the category today: Devin, OpenHands, and SWE-agent.
Devin
Devin, from Cognition, is the commercial flagship - the tool that made “AI software engineer” a mainstream phrase. It runs in a cloud environment with a full dev setup: shell, editor, browser, and planning tools. You assign work through a chat-like interface (or Slack/Jira integrations), and Devin executes, reporting progress and opening PRs.
What it’s good at:
- Well-scoped tickets - dependency upgrades, bug fixes with repro steps, small features with clear acceptance criteria
- Long-running work - it can grind on a task for extended sessions without losing context
- Team integration - it fits into existing workflows rather than asking you to change them
Where it falls short: ambiguous or design-heavy work. Devin needs the same thing a junior engineer needs - a clear definition of done. Pricing is also team-oriented rather than hobbyist-friendly.
OpenHands
OpenHands (formerly OpenDevin) is the open-source answer. It replicates the Devin concept - sandboxed agent with shell, editor, and browser access - as an MIT-licensed project you can run yourself with your choice of model.
docker pull docker.all-hands.dev/all-hands-ai/runtime:latestWhy it matters:
- Model freedom. Run it with Claude, GPT, or a local model. Your code never has to leave your infrastructure.
- Transparency. Every action the agent takes is inspectable, and the framework is a popular research testbed.
- Cost control. You pay model costs only - no platform fee.
The trade-off is polish: you’re operating infrastructure, and setup is more involved than signing into a SaaS.
SWE-agent
SWE-agent comes out of Princeton’s research group and is built around a key insight: the interface between the model and the environment matters as much as the model. Its Agent-Computer Interface (ACI) research shaped how the whole industry thinks about giving agents shell and editor access.
It’s primarily a research and benchmarking tool - the way to run SWE-bench evaluations or experiment with agent architectures - but it’s also a capable agent for repository-level bug fixing.
How They Compare
| Devin | OpenHands | SWE-agent | |
|---|---|---|---|
| Type | Commercial SaaS | Open source | Open source (research) |
| Setup | Sign up | Self-host | Self-host |
| Best for | Teams delegating tickets | Custom/self-managed agents | Benchmarking & research |
| Model | Cognition’s stack | Your choice | Your choice |
Setting Expectations
None of these replace engineers. The realistic framing: they’re tireless junior-to-mid contributors for well-specified work, and the teams getting the most value treat them that way - clear tickets, real code review on every PR, and no deployment without human sign-off.
The category is moving faster than any other in AI tooling. Track all three (plus newer entrants like Jules) in the directory.