A terminal coding agent only works as long as you're watching it, which caps how much you can hand off at however many hours you're willing to babysit a session. Background agents remove that ceiling: you hand over a scoped ticket, it runs on a cloud VM with your repo already cloned, and you come back later to a pull request. I compared the current field in more depth on DevToolLab; here's the short version.
The pricing gap between these products is bigger than the marketing implies. Google's Jules gives every account 15 free tasks per rolling 24 hours, Devin's Pro plan starts at $20/month, and OpenAI's Codex cloud rides on whatever ChatGPT plan you already have, including the free one. The Jules CLI's own --help output documents jules new --repo torvalds/linux --parallel 3 "write unit tests" as an example, meaning three cloud VMs can race the same task for free.
What actually counts as a background agent
The term gets applied loosely, so the definition matters. A background agent clones your repo into infrastructure you don't manage, works with nobody in the loop, and hands back a branch or PR. That's different from a terminal agent like Claude Code or Codex CLI running in your own session, and different again from an IDE autocomplete assistant.
That definition has consequences. You need a repo config file the agent can actually read, because there's no human around to ask clarifying questions mid-run. You need a test suite that passes clean, because CI is the only feedback loop it gets. And you need a review habit, since every output is a diff from something that never asked "wait, did you mean X or Y?"
Devin: most autonomous, easiest to overspend on
Devin defined this category, and its docs set expectations bluntly: if you could finish it in three hours, Devin can probably finish it too. Each session gets a full VM (shell, embedded IDE, browser) that you can watch or take over anytime, and work comes in from Slack, Linear, the API or Devin's own CLI.
Pricing moved away from the old per-unit model: Free with limited usage, Pro at $20/month, Max at $200/month for a bigger weekly quota with no daily cap, and Teams from $80/month with $40 full seats plus free flex seats drawing from a shared credit pool. Credits don't expire, and sessions stop billing while idle waiting on your reply or a test run. There's no cap on concurrent sessions, which the docs explicitly encourage you to exploit by splitting big projects into smaller ones.
Codex cloud: free if you already pay for ChatGPT
Codex isn't one product anymore, it's a surface spanning a CLI, an IDE extension, iOS, and the cloud piece that runs tasks in isolated environments, startable from the web, GitHub, GitLab, Linear or Slack. There's no separate subscription: it rides ChatGPT Free, Go ($8/month), Plus ($20), Pro (from $100), plus Business/Edu/Enterprise. Plus unlocks cloud integrations like automatic code review; Pro adds 5x-20x rate limits and an early-access faster model.
The local half is genuinely open, Apache 2.0 with 118,000+ GitHub stars, and running npx @openai/codex@latest --version printed codex-cli 0.149.1 when I checked. That's the appeal for teams running the same agent locally on ambiguous work and in the cloud on mechanical tickets.
Jules: the most generous free tier, with a catch
Jules clones your repo into a Google Cloud VM, plans, executes, and opens a PR. Free tier gives 15 tasks per rolling 24 hours with 3 concurrent, on Gemini 2.5 Pro; Jules in Pro bumps that to 100 daily/15 concurrent, Ultra to 300 daily/60 concurrent, both on newer models.
The catch is billing. Paid tiers ride the Google AI Pro ($19.99/month) and Ultra (from $99.99/month) consumer plans, and Google's own docs say those are currently gmail.com-only, no Workspace upgrade button, just a waitlist form. Task limits are per user too, not pooled across a team, which matters if you were hoping to share one plan.
Also worth knowing: Copilot's cloud agent and OpenHands
If your backlog already lives in GitHub issues, GitHub Copilot's cloud agent assigns from an issue or an @copilot mention, runs in an ephemeral GitHub Actions environment, and opens exactly one PR per task. It comes with any paid Copilot plan and burns Actions minutes plus AI credits instead of a new bill, but it's capped at a hard 59 minutes per session and one repo per task.
OpenHands is the self-hosted answer, MIT licensed with roughly 85,000 stars, and it's grown past being just one agent: v1.15.0 (August 21, 2026) ships Agent Canvas, a control center that can run OpenHands itself, Claude Code, Codex or Gemini across local, Docker, VM or cloud backends. Its automations feature, scheduled or webhook-triggered workflows wired into Slack, GitHub and Linear, is what turns it into a real background agent instead of a local chat window. Run the backend on a small box and it keeps working with your laptop closed, same value proposition as the paid products minus the subscription, plus the sandboxing responsibility.
Piloting one this week
Write the repo config first (AGENTS.md or equivalent), since every agent here reads one at the repo root and it's the single biggest quality lever you control. Then start free from a terminal:
npx --yes @google/jules@latest --help
npx --yes @openai/codex@latest --version
Pick something mechanical, like backfilling tests for one module, fan it out in parallel to see variance across attempts, then measure how long a human spends turning the output into something mergeable, not how long the agent ran. If that review time beats writing it yourself, you've found a task worth delegating. Only then does the pricing question have a real answer, because usage-based plans and credit-based plans behave very differently depending on how chatty your sessions run. The full DevToolLab guide has the exact commands for fanning out and pulling a finished patch back with the Jules CLI.
How to choose
Already paying for ChatGPT: Codex cloud costs nothing extra and the same agent works locally through its open CLI. Want maximum autonomy on gnarly tickets: Devin at $20 Pro, respecting its own three-hour rule of thumb. Testing the pattern on your own dime: Jules, free, 15 tasks a day, just confirm your account type can upgrade before building on it. Workflow already lives in GitHub issues: Copilot's cloud agent, as long as 59 minutes and one repo fit the job. Code can't leave your infrastructure: OpenHands, self-hosted, MIT.
Conclusion
This category's economics flipped in 2026. What used to be a premium product now has a free entry from Google, a $20 plan a lot of people already own, or an MIT-licensed stack you can run yourself. Agent time is no longer the scarce resource, review capacity is. Write the config, spend one free run on a boring ticket, and time the review instead of the run.
References
- Best Background Coding Agents in 2026: Devin, Codex Cloud and Jules - original article on DevToolLab
- AI IDE Config Generator - draft the AGENTS.md or .cursorrules file every background agent reads before it starts
- LLM Token Cost Calculator - estimate what a usage-based agent plan actually costs per task
- Devin
- Codex cloud
- Jules
- GitHub Copilot cloud agent
- OpenHands
Top comments (0)