We tested the autonomous, background AI coding agents, OpenHands, AWS Kiro, Factory AI, Devin, and Google Jules, on independent results versus vendor claims, not marketing.
An AI coding agent is a step more autonomous than the AI coding assistants on our main hub: it takes a ticket, plans multi-step work across a whole repository, runs its own tests, and opens a pull request in an isolated cloud sandbox, largely without a human driving an IDE session in real time. We scored the five most-discussed agent-native tools of 2026, OpenHands, AWS Kiro, Factory AI, Devin, and Google Jules, weighing independent test results and re-verified benchmark claims more heavily than vendor marketing. OpenHands leads at 6.7/10 (Good), the open-source option scoring ahead of the better-funded, more heavily marketed alternatives. AWS Kiro (6.2/10) and Factory AI (6.1/10) follow as Fair, with Devin (5.8/10) and Google Jules (5.7/10) rounding out the ranking, both held back by real, documented gaps between vendor claims and independently verified results.
| # | Tool | Type | Score | Tier | From | Free | Link |
|---|---|---|---|---|---|---|---|
| 01 | OpenHands | FREEMIUM | 6.7 | GOOD · T2 | $0/mo [liveFacts] ✓Jul'26 | Yes | Visit ↗ |
| 02 | AWS Kiro | FREEMIUM | 6.2 | FAIR · T3 | $0/mo [liveFacts] ✓Jul'26 | Yes | Visit ↗ |
| 03 | Factory AI (Droid) | PAID | 6.1 | FAIR · T3 | $20/mo [liveFacts] ✓Jul'26 | No | Visit ↗ |
| 04 | Devin by Cognition | FREEMIUM | 5.8 | FAIR · T3 | $0/mo [liveFacts] ✓Jul'26 | Yes | Visit ↗ |
| 05 | Google Jules | FREEMIUM | 5.7 | FAIR · T3 | $0/mo [liveFacts] ✓Jul'26 | Yes | Visit ↗ |
Open-source AI coding agent for autonomous software development
From $0/mo — check billing term.
AWS's spec-driven agentic coding IDE, built on VS Code
How we score — every tool runs the same pipeline before a number is published.
Core agent is free and MIT-licensed with no per-seat cost for self-hosting
Ranks #1 by overall score. Core agent is free and MIT-licensed with no per-seat cost for self-hosting; Model-agnostic; not locked into a single LLM provider.
Structured requirements/design/tasks workflow produces reviewable, version-controlled artifacts instead of a single opaque diff
Ranks #2 by overall score. Structured requirements/design/tasks workflow produces reviewable, version-controlled artifacts instead of a single opaque diff; Built on Code OSS, so existing VS Code settings, keybindings, and Open VSX extensions carry over.
Bundles managed access to Claude, GPT, Gemini, and other models in a single subscription instead of requiring separate vendor accounts
Ranks #3 by overall score. Bundles managed access to Claude, GPT, Gemini, and other models in a single subscription instead of requiring separate vendor accounts; Missions feature supports long, multi-feature autonomous projects with a defined plan, execution, and validation structure.
Substantial reported enterprise traction: Cognition names Citi, Mercedes-Benz, Goldman Sachs, and Santander as customers and reported a $492 million annualized revenue run-rate in May 2026 (vendor-reported)
Ranks #4 by overall score. Substantial reported enterprise traction: Cognition names Citi, Mercedes-Benz, Goldman Sachs, and Santander as customers and reported a $492 million annualized revenue run-rate in May 2026 (vendor-reported); SOC 2 Type II certified since September 2024, with an ISO/IEC 27001:2022 certification also listed on Cognition's Trust Center.
Free tier available with no waitlist
Ranks #5 by overall score. Free tier available with no waitlist; Plan-then-execute workflow lets you review intended changes before they run.
Each tool earns a 0–10 score from six weighted dimensions, then a documented editorial adjustment for risks the formula under-weights. No paid placement — affiliate links never move a score. Read the full methodology →
The tools on this page read a ticket or prompt, plan multi-step work, edit files across a repository, run tests, and open a pull request largely unsupervised, often in a cloud VM you never see. Claude Code, Cursor, and GitHub Copilot (all ranked on our main coding assistants hub) increasingly offer a similar cloud or background agent mode, and GitHub's own Copilot Coding Agent works this way against a GitHub Issue. We cover those under the assistants hub since that is still their primary interface; this page ranks tools built agent-first.
This category has an unusually large gap between marketing and outside verification. Devin's March 2024 launch demo drew a substantiated controversy over how a task was framed, and its original SWE-bench comparison measured a 25% subsample against baseline scores computed on the full dataset; the only independent real-world test we found (Answer.AI, January 2025) recorded 3 of 20 tasks succeeding. OpenHands, by contrast, has a reproducible, third-party-citable score on SWE-bench Verified using an openly documented agent scaffold, which is a meaningfully different evidence bar. Read every "X% success rate" claim in this category as a vendor claim until you see who ran the test.
OpenHands is the clear pick if you want to see and control exactly what the agent does: it is MIT-licensed and self-hostable for free, with paid SaaS and Enterprise tiers on top. Devin, Google Jules, AWS Kiro, and Factory AI are all closed, fully managed cloud products where you cannot inspect the underlying agent loop. If running an agent against your own infrastructure, with your own model choice, matters to you, that difference matters more than any headline benchmark number.
AWS Kiro stands out structurally: instead of jumping straight from a prompt to code, it first generates a requirements document, a design document, and a task list (in EARS notation) for you to review and approve before any code is written. This is a deliberate trade-off, slower to a first result than a prompt-and-go agent, but it produces an explicit, reviewable plan you can catch problems in before the agent starts editing files. Kiro also arrives as AWS is winding down its older Amazon Q Developer product, with new Q Developer signups already closed and full end-of-support set for April 30, 2027.
Devin and Factory AI both report substantial funding and named enterprise customers (Devin: Citi, Mercedes-Benz, Goldman Sachs, Santander; Factory AI: Nvidia, Adobe, EY, Palo Alto Networks, Adyen, all vendor-reported), and both bill on a usage-based credit model where real monthly costs can run well above the advertised entry price. Google Jules is the newest to reach general availability (April 2026) and publishes the most straightforward per-day task quotas of the five. None of the five publish an independently confirmed, apples-to-apples cost-per-completed-task figure, so budget for real variance above the sticker price on any of them.
Google Antigravity, a separate, newer Google product from Jules, launched in May 2026 with multi-agent orchestration and real browser-based UI verification; it is too new for a confident verdict and is worth a watch rather than a purchase decision today. Goose (from Block, now under the Agentic AI Foundation) and Crush (Charm's fork of the OpenCode project) are open-source, enthusiast-oriented terminal agents with smaller but genuine followings; neither is coding-specific enough (Goose) or broadly adopted enough (Crush) yet to rank here. Replit Agent, covered on our AI app builders roundup, works similarly to the agents here but is built for a broader, less technical audience building whole apps from scratch rather than working tickets against an existing codebase. Windsurf, once a standalone product, was acquired by Cognition in July 2025 and folded into Devin Desktop in June 2026; it is no longer a separate product to evaluate.
OpenHands ranks first in our testing at 6.7/10 (Good), the open-source option outscoring better-funded, more heavily marketed rivals thanks to a reproducible, independently citable benchmark result and full transparency into how it works. AWS Kiro (6.2/10) and Factory AI (6.1/10) follow as solid Fair-tier options.
That was Cognition's own marketing language at its March 2024 launch, not an independently verified fact. The launch demo drew a substantiated credibility controversy, and the only independent real-world test we found recorded a low task-success rate. Devin has real enterprise traction and funding since then, but treat capability claims as vendor-reported until you pilot it on your own tasks.
An agent plans and executes multi-step work across a repository largely unsupervised, running its own tests and opening its own pull request, often in a cloud sandbox. An assistant like Claude Code, Cursor, or GitHub Copilot is more actively steered by a developer in real time, even though several assistants now also offer a background or cloud agent mode. See our AI coding assistants hub for that category.
AWS is retiring it: new signups stopped on May 15, 2026, with full end-of-support set for April 30, 2027. AWS Kiro, a newer, spec-driven agent built on Visual Studio Code, is positioned as its replacement.
Yes. OpenHands is MIT-licensed and free to self-host, with paid SaaS and Enterprise tiers available if you want a managed version. It is the only tool in this ranking you can fully inspect and run on your own infrastructure.
Each tool is scored 0-10 across six weighted dimensions: capability (0.30), value (0.20), ease (0.15), privacy (0.15), support (0.10), and ecosystem (0.10), using official pricing, independent test results where available, security disclosures, and sourced developer sentiment. We weighed independently reproducible results well above vendor-only claims, with editorial overrides documented where a credibility gap or a major real-world signal was not otherwise captured by the formula. See our methodology for the full breakdown.
OpenHands is the strongest overall pick in this category precisely because its results are independently checkable and its code is open, not because it has the biggest funding round or the loudest launch. AWS Kiro is worth a serious look if a spec-first, review-before-code workflow fits how your team works, and Factory AI is the strongest option if you need genuine enterprise governance across a fleet of coordinated agents. Devin and Google Jules both have real capability, but weigh the gap between what each vendor claims and what has actually been independently verified before betting a workflow on either one.