Agents with their own computer.
Some agents work in your editor or terminal. Others get a computer of their own in the cloud, with a browser that stays signed in, files that persist and work that carries on while you're away. This page ranks the second kind on what they can do, from a checklist where every answer links to the vendor's own documentation.
Top score
100
Grok Bot
Cited answers
176
of 210; the rest are Unknown
Weights
Equal
One point per capability, ½ for partial
Its own computer, or yours?
Both kinds are excellent at different things. The difference is where the agent works and what it can reach.
Agents with their own computer
Grok Bot, Manus, Devin, ChatGPT dots and the rest of this ranking
Keeps working while you're away
Long jobs, schedules and event triggers run on the agent's machine, with your laptop closed.
Stays signed in
A persistent browser keeps its logins, so it can use the web apps you'd use yourself.
Carries a job end to end
It can build, run, check in a browser and, with some products, publish, without handing the files back.
Remembers its setup
Installed tools, files and memory persist, so the next task starts where the last one stopped.
Agents in your editor or terminal
Claude Code, Codex CLI, Cursor, Grok Build, Gemini CLI, Cline
Best-measured at code
Terminal agents top Terminal-Bench, the public board for hard terminal tasks.
Your real environment
It uses your repo, your toolchain and your credentials exactly as they are.
Tight review loop
You watch each change land and can stop or redirect it at once.
Nothing leaves your machine
Code and secrets stay local, apart from what is sent to the model.
When to pick which
- The job runs for hours, on a schedule, or while you sleepOwn computer
- It needs logged-in websites, web apps or a desktopOwn computer
- You want it built, deployed and checked without doing the last steps yourselfOwn computer
- Deep changes in a large codebase you're actively working onEditor or terminal
- Code or credentials that must stay on your machineEditor or terminal
- Short, interactive edits where you review every diffEditor or terminal
Many people use both: an editor or terminal agent to write the code, and an agent with its own computer to run, deploy and keep an eye on it.
Who can do the most on their own
Scored on 14 capabilities, equally weighted. Switch to the editor and terminal agents to compare the other kind. Ties share a rank and are listed alphabetically.
- YesYes · 1 point
- PartialPartial · ½ point
- NoNo · 0 points
- UnknownUnknown · 0 points
- UnverifiedUnverified · 0 points
Open a row for each answer, its note and the vendor page it comes from. Every icon links to its source. Scroll the table sideways on small screens.
What the public boards say
Real leaderboard results only. Nothing here is estimated.
| Benchmark | Measures | Ranked products listed |
|---|---|---|
| OSWorld-Verified | Real desktop tasks on Ubuntu, Windows and macOS | Not listed |
| GAIA | General assistant tasks needing browsing and tools | Not listed |
| BrowseComp | Hard-to-find facts on the web | Not listed |
| WebArena | Tasks on realistic self-hosted websites | Not listed |
| Terminal-Bench 4.0 | Hard tasks in a real terminal | Codex Cloud: #3, 58.18% (as the Codex CLI harness) |
| τ-bench | Tool use with simulated customers and policies | Not listed |
As of October 20, 2026, none of the ranked products appears as a product on OSWorld-Verified, GAIA, BrowseComp, WebArena or τ-bench; those boards rank models and research agents, not finished products. Terminal-Bench lists agent harnesses, so it covers the editor and terminal agents (switch the table above to see them). That's why this ranking is built on documented capabilities instead.
The checklist
Each capability is worth one point, a partial answer half a point. No, Unknown and Unverified score zero. The score is the share of the maximum, from 0 to 100. Price is shown but not scored.
- 01
Own persistent computer
Does the agent get its own cloud computer or VM that stays around between tasks?
Partial: A cloud environment that keeps some state, or one created fresh per task from a saved setup.
- 02
Files persist across sessions
Do files the agent creates stay available for later tasks?
Partial: Only a saved setup or project files persist, not the agent's working files.
- 03
Browser with saved logins
Can it use a real browser and stay signed in to sites between tasks?
Partial: It drives a browser on your own computer, or logins don't persist.
- 04
Full desktop and GUI apps
Can it operate a full graphical desktop and apps, not just a browser?
Partial: Only on your own computer, or only a browser view.
- 05
Shell and installs tools
Can it run terminal commands and install software?
Partial: Runs code, but a shell or installs aren't documented.
- 06
Deploys to hosting
Can it publish or deploy what it builds to a live URL?
Partial: Only a preview or a third-party handoff.
- 07
Scheduled and recurring tasks
Can you set work to run on a schedule?
Partial: Limited cadence or only in some surfaces.
- 08
Event triggers
Can an outside event (email, Slack, GitHub, webhook) start work?
Partial: Polling check-ins or a narrow set of events.
- 09
Long-term memory
Does it remember preferences and context across sessions?
Partial: Scoped to a project, or only through files you maintain.
- 10
Connectors and integrations
Can it connect to your apps (MCP, plugins or built-in connectors)?
Partial: A few fixed integrations only.
- 11
Works while you're away
Does work continue with your laptop closed?
Partial: Only some task types continue.
- 12
Parallel agents
Can it run several tasks or sub-agents at the same time?
Partial: Limited parallelism.
- 13
Approvals and safety controls
Are there documented approval steps or permission rules for risky actions?
Partial: Some controls, but no per-action approvals documented.
- 14
Desktop and mobile apps
Can you reach it from native desktop and mobile apps?
Partial: Only one of the two (or web only on one side).
- Yes Yes: 1 point
- Partial Partial: ½ point
- No No: 0 points
- Unknown Unknown: the vendor's public pages don't say
- Unverified Unverified: Grok Bot only: not shown publicly by xAI
Left out, and why
- ChatGPT agent (agent mode)
Retired. OpenAI says to use ChatGPT Work instead, which is ranked.
- Project Mariner
Shut down on May 4, 2026. Google points users to Gemini Spark, which is ranked.
- Claude in Chrome
A browser extension that works in your own Chrome. Its abilities are counted in the Claude (Cowork) row.
- Perplexity Comet
A browser on your computer. Perplexity's cloud agent is Computer, which is ranked.
- Genspark Super Agent
Session-based sandbox without schedules. Genspark's persistent product is Claw, which is ranked.