What can AI agents actually do?

The core capability: act, don't just answer

The defining thing an AI agent can do is take an action in the world. A language model on its own predicts text; an agent wraps that model in a loop and a set of tools so it can do something with each thought. Give it a browser and it can navigate sites and fill forms; give it a shell and it can run code; give it an API and it can update a CRM or send an email. The model decides what to do, the tools carry it out, and the agent reads the result to decide its next step.

That loop — perceive, decide, act, repeat — is why agents can finish multi-step tasks that a single prompt cannot. A chatbot can tell you how to migrate a database; an agent can do the migration, run the tests, and hand you a pull request. The capability is bounded by the tools and permissions you grant, which is both the source of an agent's power and the thing you have to manage carefully.

Where agents are genuinely good today

Coding is the standout. Agents like Claude Code, Devin, and Cursor read entire codebases, make multi-file changes, run tests, fix their own errors, and open reviewable pull requests — and they're measured on real benchmarks like SWE-bench, so the progress is verifiable rather than hype. Research is a close second: tools like Elicit and GPT Researcher plan searches, read dozens of sources, and produce structured, cited reports in minutes.

Beyond those, customer-support agents such as Intercom's Fin resolve tickets end to end and are priced per resolution because they actually close cases. Sales agents like Clay and Ava handle prospecting and outreach. Website builders like Lovable and v0 turn a description into a deployed app. General agents like Manus run long tasks asynchronously and return finished deliverables. Each of these is in production, not a demo.

What agents still can't do well

Agents are reliable in proportion to how clearly a task is defined. They excel when there's an unambiguous goal and a way to check the work — code that passes tests, a form that submits, a report that cites its sources. They struggle with vague, open-ended, or high-judgment work where 'done' is subjective, where the right approach is itself the hard part, or where a mistake is costly and irreversible.

They also make confident errors. An agent can hallucinate a fact, follow a malicious instruction hidden in a web page, or take a wrong action with full conviction. That's why the well-run deployments keep a human in the loop on anything irreversible and scope each agent's access to the task at hand. The honest summary: agents can do a remarkable amount of real work, but they're powerful assistants to supervise, not autonomous employees to forget about.

What can AI agents do that chatbots can't?

Agents take actions, not just produce text. By using tools — a browser, a shell, an API — in a loop, an agent can complete a multi-step task end to end (edit code and open a pull request, resolve a support ticket, book a meeting), where a chatbot can only describe how to do it.

What is the most useful thing AI agents do today?

Coding is the most proven use. Agents read a codebase, make changes, run tests, and open reviewable pull requests, and their progress is measured on real benchmarks. Research, customer support, sales, and website building are other domains where agents do genuine production work.

Can an AI agent run a task completely on its own?

For bounded, well-defined tasks with a clear way to check the result, largely yes. For vague, open-ended, or high-stakes work, agents need supervision — well-run deployments keep a human approving anything irreversible rather than letting the agent act unchecked.

What can't AI agents do reliably?

They struggle with ambiguous or high-judgment tasks where 'done' is subjective, and they can make confident mistakes or follow malicious instructions hidden in content they read. They're powerful assistants to supervise, not autonomous workers to leave unattended.