Autonomous agents: what they are, how they fail and when to skip them

The short version

• An autonomous agent chooses its own next step towards a goal, rather than following a path written in advance.
• Autonomy is a matter of degree, set action by action: the useful question is what an agent may do without a person.
• Agent errors compound, because an early mistake feeds every later decision in the run.
• An agent's permissions should be enforced in code outside the model, because a model can be misled by what it reads.
• If the steps are known in advance, ordinary workflow automation is cheaper and easier to test than an agent.

What is an autonomous agent?

An autonomous agent is software that pursues a goal by choosing its own steps, acting through tools and deciding when it has finished. You give it an outcome rather than a procedure. In current systems a language model usually does the choosing, and ordinary code carries out the actions.

The defining feature is the loop. The agent reads the situation, picks an action, takes it, reads the result and decides again. A system that answers once and stops is not an agent, however good the answer. A system that runs a sequence somebody wrote in advance is automation, even if a model fills in one of the steps.

Autonomy is a matter of degree. An agent that drafts a refund for approval and an agent that sends the money are both agents. They differ in what they may do without a person, and that is a design decision, not a property of the model.

How does an autonomous agent work?

An autonomous agent works by running a model in a loop with access to tools, until the goal is met or a limit is reached. Four parts matter: the model, the tools, the state and the harness, which is the code around the model.

The harness gets the least attention and often decides whether the agent works. The Center for AI Safety reports that part of the recent rise on its Remote Labor Index comes from stronger agent scaffolding, not only newer models. A better model in a poor harness can still lose track, repeat itself or stop early.

What are the parts of an autonomous agent?

• •The model reads the goal and the current state, then picks the next action.
• •Tools are the functions it can call, such as search, read a record, send a message or run code.
• •State is what it carries between steps: the goal, what it has tried and what came back.
• •The harness runs the tools, checks permissions, counts steps and stops the run.

What is an example of an autonomous agent?

A support agent handling a refund request is a simple example of an autonomous agent. Given 'refund this order if it qualifies', it looks up the order, reads the refund policy, checks the delivery record, and either issues the refund or passes the case to a person with its reasoning attached.

Nobody wrote that path in advance. If the delivery record is missing, the agent can look for a courier note or ask the customer a question. That flexibility is the reason to use an agent. It is also why an agent is harder to test, because the same request can take a different route on a different day.

Coding agents are another familiar example. They read a repository, edit files, run the tests, read the failures and try again.

How is an autonomous agent different from a chatbot or workflow automation?

The difference is who decides the next step. A chatbot replies and waits for a person. Workflow automation follows steps a person wrote in advance. An autonomous agent decides the next step itself at run time, which gives it flexibility and costs it predictability.

How do chatbots, workflow automation and autonomous agents compare?

How reliable are autonomous agents today?

Autonomous agents are improving quickly on benchmarks, but no benchmark tells you how one will do on your process. The published measures are narrow, and each has limits you should read before quoting it.

METR's Time Horizon 1.1, published on 29 January 2026, tests frontier models on 228 tasks, mostly software and research-engineering work. METR reports a doubling time of 89 days on data since 2024. The suite does not measure general business tasks, and its long trend uses earlier-suite estimates for pre-2023 models.

The Remote Labor Index is closer to ordinary paid work. Agents attempt real freelance projects, and human evaluators judge each deliverable against a gold-standard one from a paid professional. Scale Labs' live leaderboard showed a top score of 21.25 when read on 6 November 2026. That figure will move, so check the leaderboard and its definition before repeating it.

Benchmarks only partly show how errors compound. A step that is right almost every time, repeated across a long run, still produces runs that go wrong. An early mistake also feeds every later decision. Longer tasks need more checkpoints, not only better models.

When should you use an autonomous agent?

Use an autonomous agent when the steps vary from case to case, the result can be checked, and a mistake is cheap or reversible. All three should hold. If one is missing, a simpler system or a person approving each action is usually the better design.

Good candidates tend to be gathering and preparing work: researching a supplier, triaging a queue, assembling a case file, drafting a reply for someone to send, or proposing a code change for a person to review. The agent finds the route and a person keeps the decision.

When is an autonomous agent not worth it?

An autonomous agent is not worth it when the steps are known in advance. A fixed workflow will be cheaper to run, easier to test and easier to explain when someone asks why something happened. Adding a model to choose steps that never change buys variation you did not want.

It is also a poor fit when a wrong action cannot be undone and nobody can check in time, when you cannot define a correct result, or when the volume is too low to repay the build. Deciding not to build one is a sound outcome. Plenty of 'agent' projects are workflow projects with a more exciting name.

What are the risks of autonomous agents?

The main risks of autonomous agents are wrong actions taken with real permissions, instructions hidden in content the agent reads, costs that grow unnoticed, and failures that look like success. They share a cause: the agent acts faster than anyone is watching.

Hidden instructions need particular care. An agent reads emails, web pages and documents, and a model cannot reliably separate the content it is reading from the instructions it should obey. If the agent that reads a stranger's email can also send money, that email is a route to your money. Enforce permissions in code the model cannot talk its way past, and assume anything it reads may be hostile.

Controls can also arrive after the building starts. Pax8's own Q3 2026 online survey covered 1,150 owners, executives, functional leaders and managers at companies with 5 to 499 employees in the US, UK, Germany and Sweden. Its headline is that small businesses are building AI faster than they can govern it. It is a vendor's survey, so treat it as a signal rather than a measurement.

How much autonomy should an autonomous agent have?

Give an autonomous agent the least autonomy that gets the job done, then widen it for specific actions once a record shows it has earned that. Decide per action, not per agent. The same agent might search freely, draft freely and still need approval to send anything.

What are the levels of agent autonomy?

1. Suggest: the agent recommends, and a person does the work.
2. Draft: the agent prepares the action, and a person approves each one.
3. Act and report: the agent takes reversible actions, and a person reviews them afterwards.
4. Act within limits: the agent takes consequential actions inside fixed caps, and anything outside them goes to a person.

What controls should every autonomous agent have?

• •Permissions enforced in code outside the model, scoped to the person or task it is acting for.
• •Approval gates on anything that cannot be undone, such as payments, deletions and external messages.
• •Hard limits on steps, time and spend for each run.
• •A log of every tool call with its inputs and result.
• •A test set of real cases, re-run whenever the model, the prompt or a tool changes.
• •A way to stop a run and switch the agent off.

What do autonomous agents cost?

Autonomous agents cost a variable amount per task, because the number of steps changes from run to run. Budget per completed task, including failed runs and human review, rather than per model call. We do not have a reliable general figure for total cost, and one given without seeing the process is a guess.

Published platform pricing gives one reference point. Salesforce's Agentforce pricing page, read on 6 November 2026, is based on $500 per 100k Flex Credits and also lists $2 per conversation. The page says pricing is subject to change. Neither figure covers your own integration work, testing or the people checking outputs.

How widely are autonomous agents used?

We do not have a reliable figure for autonomous agent use, because the best official source we read does not separate agents from other AI. The US Census Bureau's Business Trends and Outlook Survey is biweekly, samples about 1.2 million businesses and represents US employer businesses excluding farms. Its story covering 14 December 2025 to 3 May 2026 reports firms with at least 20 employees as the biggest AI users. It does not break out AI agents.

So be careful with confident adoption numbers for agents. Check who published the figure, who was asked and whether it separates agents from AI in general. The word 'agent' is used loosely enough that two surveys may not be counting the same thing.