What Can AI Agents Actually Do? What Benchmarks Show

AI agents do best on bounded, checkable work: tasks with clear inputs, tools they can use and a result that can be verified. They are less dependable on long, cross-application work with vague goals, where mistakes are expensive and hard to spot.

We describe what benchmarks test rather than quote their scores, because results change weekly. See current positions on the leaderboard and how we combine them in our methodology. You can also read the knowledge work board and coding board for your own job type.

Disclosure: eSEOspace is a member of the Claude Partner Network.

At a glance

These fit levels are our editorial judgment, drawing on what the benchmarks below test. They are not measured results.

Benchmarks test more than you might think

An AI agent is a model plus tools and a loop that lets it act. So agent benchmarks need realistic work, not exam questions. Here is what the main ones actually measure.

That is already a useful finding. Serious people now test agents on professional work, not only on puzzles.

What agents can do reliably

Work with fast feedback

Coding is the clearest case. A test suite tells the agent, within seconds, whether it succeeded. So it can try, fail, read the error and fix it.

Your other workflows can borrow this idea. Anything with a checker, such as a validation rule or a reconciliation total, tends to suit agents better. Anthropic's evaluation guide points to well-specified tasks, stable environments and thorough tests for coding agents.

Bounded tool use

Agents can call tools, read results and take the next step. Vals lists tool use among the capabilities its benchmarks test, along with reasoning, long context and long-horizon tasks.

The tighter the task, the better this works. "Look up this order and draft a reply" suits an agent well. "Improve our customer experience" does not.

Reading long documents

Models can now hold a lot of text at once. For example, Google lists a 1M token context window on Gemini 4 Argon as of 2 November 2026. That helps with contracts, codebases and reports.

Capacity is not the same as accuracy. An agent can read the whole contract and still miss the clause that matters.

Where agents still struggle

Long tasks

Reliability falls as tasks grow longer, because one early mistake can spoil everything after it. METR defines a time horizon as the task length, measured by human expert time, at which an agent succeeds at a given rate.

Its two lines, for 50% and 80% success, show the point well. A task an agent handles half the time is not one you can leave unattended. METR's page says it is no longer actively updated, so treat it as a concept and check dates.

Cross-application professional work

APEX-Agents asks agents to work through files and tools on tasks written by analysts, consultants and lawyers. It scores with Pass@1, meaning the first attempt must pass. That is close to how a client experiences your work.

The authors of GDPval also note that more reasoning effort, more task context and more scaffolding improve results. So setup matters as much as the model. Expect to spend real time on instructions and context.

Judging their own success

Agents can report success when the work is wrong, or find shortcuts. Terminal-Bench's integrity update defines reward hacking as exploiting a loophole without showing the skill being tested. Its usual example is an agent finding solutions online.

Terminal-Bench plans tamper-resistant verifiers in a future release, and that tells you how much this matters. In your own use, never let an agent be the only judge of its own work.

Knowing when to stop

Terminal-Bench's 4.0 notes say the remaining errors are largely model refusals and output limits. It also notes large variance in how long runs take on some models. So budget and time can be unpredictable, even on a good run.

Resisting manipulation

Agents read content they cannot fully trust. Anthropic says no browser agent is immune to prompt injection, where hidden text gives an agent instructions. Our security guide covers this in full.

Do agents actually save time?

Benchmark success does not automatically mean time saved. METR's early-2025 trial found experienced open-source developers took 19% longer with AI tools, even though they expected a speedup.

METR now says those results no longer reflect current tools. Its February 2026 update says developers are likely more sped up now. It also calls its new data only very weak evidence, because of selection effects.

Our take: be humble about productivity claims, including vendors' claims. People are poor judges of their own time savings, so measure review and rework directly.

A quick test for your own workflow

Score a candidate task with these questions. The more yes answers, the better the fit.

• Can you describe the finished result in one sentence?
• Can software or a quick human check verify the result?
• Is a mistake cheap to fix?
• Is the input reasonably clean and consistent?
• Do you have 10 to 20 past examples to test on?
• Can the agent reach the data through a narrow tool?

If most answers are no, try narrowing the task. Split a big job into small steps, and let the agent do only the checkable ones. Our human-in-the-loop guide shows where to keep people involved.

How to read a claim in a sales pitch

Vendors often quote benchmarks when they sell. A few questions will show you how much weight that number can bear.

• Which version? Terminal-Bench alone has shipped several versions, and its team changes tasks between them. Old and new results may not compare.
• Which setup? A benchmark tests a model inside a harness, the software wrapped around it. The product you buy may use a different one.
• Whose work? Finance tasks tell you little about your support queue. Ask which tasks resemble yours.
• Who funded it? Terminal-Bench says its leaderboard experiments have grant support from several AI labs, and it publishes that fact. Look for that kind of openness elsewhere.
• What did it cost? Some boards show cost and tokens next to results. A strong result at many times the cost may not suit you.

If a vendor cannot answer these, treat the claim as marketing. Then ask for a short trial on your own data instead.

Which should you choose

Choose an agent when the task is bounded, verifiable and repeated often. Coding on a branch, ticket triage and first-draft research are typical starting points.

Choose an agent with approval steps when the work is valuable but the mistakes are costly. Have it prepare, and have a person decide.

Choose a simple workflow or no AI when the steps never change, or the output can never be checked. Anthropic's guide also recommends the simplest option that works.

Then prove the fit. Use our agent scorecard on your own tasks before you commit to a budget.

Frequently asked questions