Short answer: OpenAI Codex scores highest in our provisional ranking: it reads your repo, runs tests and proposes a pull request from the CLI, IDE, web or desktop app, and it needs no separate subscription (ChatGPT Go $8/mo, Plus $20/mo; a free desktop tier is rolling out). Jules and Cursor follow on free tiers and price; Claude Code and Devin score high on quality but need a paid plan from $20/mo. Scores are editorial estimates, not benchmark results.
Which coding agent should I give a real task in my repo?
Last checked: 2026-10-18Score: provisionalDemand signal: No X count yet; added for the agent focus, 18 Oct 2026
Scores are editorial estimates from public information. Hands-on tests have not started yet. Methodology →
Scores are provisional editorial estimates: price and free-tier scores are read from each tool's official pricing page on the date shown; quality, ease and trust come from the vendors' own docs and product pages. No hands-on tests were run for this page.
What happens when a coding agent tries to fix a staging error alone
What they did
PocketOS, a small software firm serving car-rental companies, had a Cursor agent (running Claude Opus 4.6, per ABC News) hit a credential mismatch in staging. It chose on its own to fix it by deleting a Railway volume.
How it went
The volume held the production database and its backups; about three months of rental reservations vanished in roughly nine seconds. Railway restored the data about 30 minutes after the CEO got through to Railway's chief, and Railway has since patched the legacy delete endpoint.
Worth knowing
Reported by the CEO and ABC News, not independently audited. His lesson: destructive actions should need a confirmation the agent cannot give itself, and he stays "extremely bullish" on AI.
Assign bug fixes and cleanups to GitHub Copilot coding agent
What they did
Microsoft's .NET team let maintainers assign work to Copilot coding agent in the dotnet/runtime repo for ten months, then published the numbers.
How it went
Of 878 agent pull requests, 535 were merged (67.9% of those decided, versus 87.1% for Microsoft staff PRs) and 3 were later reverted. Cleanup and removal work did best (84.7%); performance work did worst (54.5%). A shared instructions file lifted success from 38.1% to 69%.
Worth knowing
Self-reported by the team that ships the product. Every PR was requested by a maintainer, 52% of merged ones needed human commits, and the author says review capacity became the bottleneck.
Fire off code research jobs to async agents and read the results later
What they did
Simon Willison ran two or three short research prompts a day on fire-and-forget coding agents (Codex Cloud, Claude Code for web, Jules, Copilot coding agent) in a throwaway GitHub repo, for example benchmarking Python Markdown libraries and compiling a C extension to WebAssembly.
How it went
About two weeks in he had 13 projects. The WebAssembly build only worked after he pushed the agent past an early give-up, and the tag-suggestion run skipped the requested visualization. He called the unreviewed reports "total slop" he would not publish unedited.
Worth knowing
He runs this on fresh, non-sensitive repos with open network access; locked-down networking is the default for private code, and an agent failing to find a solution does not prove none exists.
What is the difference between a coding agent and an AI code editor?
A coding agent takes a whole task (an issue, a bug, a feature), plans it, edits several files, runs the tests and hands you a pull request to review. An editor assistant such as autocomplete suggests code while you type. Several tools do both: Cursor and GitHub Copilot started as editors and now ship agents. For editor-first picks see our best AI for coding ranking.
Which coding agent can I try for free?
Jules lists a free tier of 15 tasks a day, Google Antigravity has a free Individual plan with weekly limits, and Cursor has a Free plan. OpenAI Codex is free inside the desktop app subject to rollout. Claude Code is not on Claude Free; Devin has a free plan with a light quota.
Does the agent merge code on its own?
No. Every tool on this list opens a pull request or shows a diff that a developer reviews and merges. Keep it that way: give the agent a branch, require tests to pass, and never let it push to main unattended.
Does anyone pay for these positions?
Rankings are built to avoid bias: no payment, commission or sponsorship changes a score or a position. Some links may become affiliate links (marked “Affiliate link”, rel="sponsored"). No affiliate links are active yet.
Weekly: the best new agents and tools + one ranked fix
