AI agents · Choosing by task

Best AI agent: pick by the kind of work, then check how much it does alone

An agent is worth picking only if it finishes the work you hand it with fewer corrections than you would make yourself. This page sorts agents into four kinds of work, gives you five checks for any of them, and shows which ZeroTwo agent each plan includes. We have not run a published head-to-head across vendors, so there is no ranking here.

Autonomy ladder · a description, not a score

  1. Level 1

    Answers

    You ask, it replies. One question at a time.

  2. Level 2

    Uses tools

    Searches the web or reads your files when you point it there.

  3. Level 3

    Runs a task

    Plans the steps, works through them and returns a result to review.

  4. Level 4

    Runs on a schedule

    Repeats a task without you starting it each time.

Four kinds of agent, four different tests

A research agent that does well on a report tells you nothing about how it would handle a codebase. Test each kind on its own task.

Research agents

Read many sources and return a cited report. The right pick when the output is a document you will verify.

Common task to run
One question with the same source constraints, such as only primary sources or only the last twelve months.
How to check it
Open every citation. Count how many say what the report claims.
In ZeroTwo
Deep research, with 150 runs a month on Plus and no stated cap on Pro and above (every run uses credits). Per the feature page, a report typically takes 5 to 30 minutes, you are notified when it is done, and it can be shared or downloaded as PDF or docX.
Deep research in ZeroTwo

Coding agents

Read a repository, plan a change, edit across files and run the tests.

Common task to run
The same small bug with a failing test you wrote.
How to check it
Run the tests yourself and read the diff.
In ZeroTwo
The code agent runs in the desktop app and is included from Plus up.
ZeroTwo code agent

Work and app-action agents

Take a goal, work across your apps and files, and stop before anything consequential.

Common task to run
A multi-step admin task with one step that needs your approval.
How to check it
Check that it paused at the approval step and did not skip it.
In ZeroTwo
The Work agent is included from Plus up.
ZeroTwo Work agent

Scheduled agents

Repeat a task without you starting it: recurring reports, monitoring, briefs.

Common task to run
A weekly summary of something you can verify yourself, run for three weeks.
How to check it
Read the run history. Do failures surface, or do they pass silently?
In ZeroTwo
Scheduled tasks run recurring work across your apps and files, with run history, outputs and approvals to review. Check the pricing page for plan availability.
ZeroTwo scheduled tasks

Five checks to run on any agent

These apply to ZeroTwo and to every other agent. The two that marketing pages leave out are the intervention count and the failure test.

Five checks for any AI agent, what to do for each, and what to write down.
CheckWhat to doWhat to write down
1Task completionGive it one task with a finish line you defined before the run.Done, partly done or not done, judged against your finish line and not against the agent's own summary.
2Human interventionsCount every time you correct it, re-prompt it or approve an action.The number, and what each intervention was about.
3Source and output qualityOpen every citation, diff or created file.How many checked out and how many did not.
4Cost of the runNote what the run used: credits, minutes or requests.Usage per completed task, not per attempt.
5Failure behaviorInterrupt it, give it a broken input, and ask for something it cannot do.Whether it stops and says so, or carries on with confidence.

Which ZeroTwo agent each plan includes

Agent access is gated by plan, and agent runs are metered: deep research draws from the same monthly credit pool as model runs, images and video.

Agent features and allowances for the Free, Plus and Pro plans, with monthly prices in the column headers.
FeatureFree$0/moPlus$14.99/moPro$29.99/mo
Deep researchLimitedYesYes
Work and Code agentsNoYesYes
Web searchSerperExaExa
Model selectionLimitedUnlimited fast modelsBetter selection
Monthly credits—1,6503,500
Free Auto requests / month—1,0002,500

Source: ZeroTwo's own plan data, the same as the table on the pricing page, where Pro 2× and Ultra raise the allowances. Heavier tasks use more credits, top-ups are available and purchased credits never expire.

What a ZeroTwo research run looks like

The steps below follow the deep research feature page. Run the five checks on the report you get back.

  1. Select deep research

    Choose Deep research from the action menu in the chat box.

  2. Assign the task

    Describe the research you want. A more detailed question gives a more usable result.

  3. Receive the report

    A report typically takes 5 to 30 minutes, and you get a notification when it is ready.

  4. Share or download

    Share a link, or download the report as a PDF or docX file.

If you need this, pick that

Autonomy costs credits and review time, so a simple task does not need an agent. If your whole workflow already lives inside one vendor's suite, that vendor's agent may fit better; check its current plan limits on its own pricing page.

For the product side, you can see supported agent execution workflows.

The model behind the agent

ZeroTwo agents run on the models your plan allows. These providers appear in our model catalog; the live list with current versions is on the models page.

  • OpenAI
  • Anthropic
  • Google Gemini
  • xAI Grok
  • Mistral
  • Perplexity
  • DeepSeek
  • Cohere
  • Qwen
  • Groq
  • Venice
  • Z.ai

Frequently asked questions

What is an AI agent?

An AI agent works toward a goal across several steps instead of answering one question at a time. It plans, uses tools such as web search or your files, and returns a result for you to review. How much it does without you is the most useful way to compare agents, which is why this page sorts them by kind of work and level of autonomy.

What is the best AI agent?

It depends on the kind of work, and we have not run a published head-to-head across vendors, so we will not name one. Pick the kind of agent that matches your task (research, coding, app actions or scheduled work), run one common task with a finish line you define in advance, and count the corrections you had to make.

Can an AI agent do research for me?

Yes, within limits. A research agent can search the web, read sources and draft a cited report, but you should open the citations and check the claims that matter. In ZeroTwo, deep research is limited on the Free plan, 150 a month on Plus and included on Pro and above, and a report typically takes 5 to 30 minutes.

Are AI agents free to use in ZeroTwo?

Partly. The Free plan needs no card and starts with 100 free credits, limited tools and limited deep research, but it does not include the Work agent or the code agent. Plus ($14.99 a month) adds Work and Code agents and access to deep research, and Pro ($29.99 a month) adds Bots, 30 free Basic cloud computer hours per month, 1,500 Firecrawl credits per month and 10,000 ElevenLabs credits per month. Heavier runs draw from a monthly credit pool.

Is deep research unlimited on Pro?

No. Deep research, model runs, images and video all draw from one monthly credit pool, and heavier tasks use more credits. Pro includes 3,500 monthly credits and 2,500 free Auto requests / month. You can top up, and purchased credits never expire.

How do I know an agent did the work correctly?

Judge it against a finish line you set before the run, open every citation, diff or created file, and count the interventions you had to make. An agent's own summary of what it did is not evidence that it did it.

Explore more

Test one agent on one real task this week, and count the corrections