
What you will read
These are the questions about AI agents that come up most on Google and in AI chats. Each answer starts short, shows what the research says, and ends with what I think.
- What is an AI agent?
- What is the difference between a chatbot and an AI agent?
- What can AI agents do today?
- What are examples of AI agents?
- How do you build an AI agent?
- Do AI agents make mistakes?
- Are AI agents safe?
- When should you not use an AI agent?
- How do you measure whether an AI agent works?
- Will AI agents replace employees?
01
What is an AI agent?
An AI agent is a system that carries out tasks on a user’s behalf, choosing its next steps and using tools rather than just answering questions. That independence requires testing and human oversight for sensitive actions or actions that are hard to undo.

Customer support and coding are among the uses Anthropic considers promising. But agents can cost more, take longer and compound errors, so its guide recommends starting with the simplest solution, which may not need an agent at all.
- Anthropic’s guide distinguishes agents, which adjust their steps based on feedback from their environment, from workflows that follow paths predefined in code.Building effective agents · Erik Schluntz, Barry Zhang (Anthropic), Anthropic, 2024
- OpenAI’s guide describes agents as systems that carry out tasks independently and recommends human oversight for sensitive, irreversible or high-stakes actions such as payments.A practical guide to building agents · OpenAI, OpenAI, 2025
What I think: I would expand an agent’s use only after measuring its results and defining when it must call a person.
02
What is the difference between a chatbot and an AI agent?
A simple chatbot answers questions; an AI agent can decide the next steps and use tools to carry out tasks on the user's behalf. The key difference is the autonomy to act, not just converse.

This distinction matters when choosing how to automate a task. According to OpenAI, agents suit work involving complex decisions, rules that are hard to maintain, or large amounts of unstructured information; otherwise, a program with fixed rules may be enough.
- OpenAI's guide distinguishes agents from simple chatbots and recommends human oversight for sensitive, irreversible, or high-stakes actions such as payments.A practical guide to building agents · OpenAI, OpenAI, 2025
- Wikipedia notes that there is no agreed definition of an AI agent but highlights the ability to make decisions and carry out multi-step tasks using external tools.AI agent · Wikipedia contributors, Wikipedia, 2026
What I think: I would measure results before giving an agent more autonomy and keep human oversight for sensitive actions.
03
What can AI agents do today?
AI agents can already perform computer tasks such as browsing the web, writing code and running programs, but they still fail frequently. In a test simulating a small software company, the best agent completed 30% of tasks on its own.
Agents’ capabilities are improving rapidly: Stanford’s 2026 AI Index reports that success on OSWorld, a test of real computer tasks, rose from 12% to about 66%. That progress still leaves roughly one in three attempts unsuccessful on the test.
- Stanford’s report records that performance on SWE-bench Verified, a coding test, rose from 60% to near 100% in one year.The 2026 AI Index Report · Stanford Institute for Human-Centered Artificial Intelligence (HAI), Stanford HAI, 2026
- METR researchers estimated that, at the time of the study, Claude 3.7 Sonnet completed software tasks taking experienced people about 50 minutes half the time.Measuring AI Ability to Complete Long Software Tasks · Thomas Kwa, Ben West, Joel Becker et al. (METR), NeurIPS 2025; arXiv:2503.14499, 2025
- A study simulating a small software company found that the best agent completed 30% of tasks on its own, while harder and longer tasks remained out of reach.TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks · Frank F. Xu, Yufan Song, Boxuan Li et al., arXiv:2412.14161, 2025
What I think: I recommend keeping pace with this progress by testing agents on the tasks they will actually handle, measuring success before promising results, and defining when they should call a person.
04
What are examples of AI agents?
Examples of AI agents include OpenAI's Deep Research and Operator, and Cursor, for research, web browsing and coding tasks. These systems can use tools and carry out multi-step tasks, but that autonomy does not guarantee correct results.
There is no agreed definition of the term. Anthropic's guide distinguishes agents, where the model directs the process and tool use, from workflows that follow predefined code paths. That difference matters when choosing a system: the company recommends starting with the simplest solution, which may not need an agent at all.
- Wikipedia lists customer-support, research, coding and web-browsing agents, with examples including Deep Research, Operator and Cursor.AI agent · Wikipedia contributors, Wikipedia, 2026
- Anthropic's guide names customer support and coding as promising uses, but warns that agents can compound errors and recommends extensive testing in isolated environments with safeguards.Building effective agents · Erik Schluntz, Barry Zhang (Anthropic), Anthropic, 2024
- In Anthropic's shop experiment with a Claude-based agent, the system found suppliers but also sold items at a loss and made up an account for receiving payments.Project Vend: Can Claude run a small shop? (And why does that matter?) · Anthropic, with Andon Labs, Anthropic, 2025
What I think: I recommend that, as agents gain autonomy to act, you test their results in an isolated environment before trusting them with real tasks.
05
How do you build an AI agent?
To build an AI agent, combine a language model, tools and instructions so it can run a task and choose which tools to use. Start with the simplest solution, test in an isolated environment and keep human oversight for sensitive, irreversible or high-stakes actions.
The choice is not just which model to use, but whether the task needs an agent. Anthropic warns that agents can increase costs, response times and the buildup of errors; OpenAI recommends them for tasks involving complex decisions, hard-to-maintain rules or large amounts of unstructured data.
- Anthropic’s guide recommends starting with the simplest solution, which may not require an agent, and conducting extensive testing in isolated environments with safeguards.Building effective agents · Erik Schluntz, Barry Zhang (Anthropic), Anthropic, 2024
- OpenAI’s guide describes an agent as a combination of model, tools and instructions and recommends human oversight for sensitive, irreversible or high-stakes actions.A practical guide to building agents · OpenAI, OpenAI, 2025
- A study on ReAct showed that alternating reasoning and actions, such as looking up external information, reduced fabricated answers and error propagation compared with reasoning without those actions.ReAct: Synergizing Reasoning and Acting in Language Models · Shunyu Yao et al., ICLR 2023, 2023
What I think: I would define the task clearly, measure errors in tests and keep human oversight for high-stakes actions before expanding the agent’s autonomy.
06
Do AI agents make mistakes?
AI agents make mistakes: they can invent information, perform tasks incorrectly, and fail when repeating a task they previously completed successfully.
Agents have improved rapidly on computer and coding tests, according to Stanford. But that progress still comes with frequent failures, and an Anthropic experiment showed mistakes with practical consequences, such as selling products at a loss and inventing an account for receiving payments.
- According to Stanford's 2026 AI Index, agents' success on OSWorld, a test of real computer tasks, rose from 12% to about 66%, but they still fail roughly one in three attempts.The 2026 AI Index Report · Stanford Institute for Human-Centered Artificial Intelligence (HAI), Stanford HAI, 2026
- Anthropic reported that an agent running a small shop sold items at a loss and invented a payment account, leading the company to conclude that it would not hire the agent.Project Vend: Can Claude run a small shop? (And why does that matter?) · Anthropic, with Andon Labs, Anthropic, 2025
- A study involving simulated customer service found success rates below 50% even among leading agents and, in retail, less than a 25% chance of completing a task successfully in all eight attempts.τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains · Shunyu Yao et al., arXiv:2406.12045, 2024
What I think: I recommend that, as agents improve, expanding their use should depend on repeated testing and clear rules for acknowledging limits and calling a person.
07
Are AI agents safe?
AI agents can leak data or take harmful actions. OWASP recommends limiting their permissions and requiring human approval for high-risk actions.
When AI can use tools, an error can become an action, not just an incorrect answer. Instructions hidden in web pages or files can also redirect its behavior, making the access granted to an agent a central safety concern.
- OWASP warns that instructions hidden in outside content can cause data leaks or tool misuse and says it is unclear whether foolproof prevention exists.LLM01:2025 Prompt Injection · OWASP Gen AI Security Project, OWASP Top 10 for LLM Applications 2025, 2025
- OWASP recommends giving agents only the tools and permissions they need and requiring human approval for high-impact actions.LLM06:2025 Excessive Agency · OWASP Gen AI Security Project, OWASP Top 10 for LLM Applications 2025, 2025
- In a study using simulated companies, Anthropic and partner researchers observed blackmail or information leaks in some tests, but found no evidence of these behaviors in real deployments.Agentic misalignment: How LLMs could be insider threats · Aengus Lynch et al., Anthropic, 2025
What I think: I would expand an agent’s autonomy only after evaluating its risks, while keeping access to a minimum and requiring human approval for high-impact actions.
08
When should you not use an AI agent?
Do not use an AI agent when a process with fixed rules already handles the task. For sensitive, irreversible or high-stakes actions, do not let the agent act without human oversight.
An agent can choose tools and carry out a task on its own, rather than just follow predefined steps. That autonomy can bring extra cost, delays and accumulating errors, so not every task justifies using one.
- Anthropic’s guide recommends starting with the simplest solution, even if that means no agent at all, and testing extensively in isolated environments with safeguards.Building effective agents · Erik Schluntz, Barry Zhang (Anthropic), Anthropic, 2024
- OpenAI’s guide recommends agents for tasks involving complex decisions, hard-to-maintain rules or large amounts of unstructured data, and calls for human oversight of sensitive actions such as payments.A practical guide to building agents · OpenAI, OpenAI, 2025
- Simon Willison recommends avoiding agents that combine access to private data, exposure to untrusted content and the ability to send information externally, because an attacker could trick them into leaking data.The lethal trifecta for AI agents: private data, untrusted content, and external communication · Simon Willison, Simon Willison's Weblog, 2025
What I think: I would start with the simplest solution and expand agent use only after measuring results and ensuring human oversight of sensitive actions.
09
How do you measure whether an AI agent works?
To measure whether an AI agent works, assess how many tasks it completes correctly, whether it repeats that success across attempts, and how much it costs. Use tasks separate from those used to tune the agent, so a good test score is not mistaken for the ability to solve problems.
Agents’ ability to complete longer tasks has been growing, but reliability and cost remain problems: one successful attempt does not show whether an agent can repeat the result or how much achieving it will cost.
- METR researchers measure capability by the time an experienced person would need for tasks an agent completes in half its attempts, and found that this duration had been doubling roughly every seven months since 2019.Measuring AI Ability to Complete Long Software Tasks · Thomas Kwa, Ben West, Joel Becker et al. (METR), NeurIPS 2025; arXiv:2503.14499, 2025
- A study using simulated customer service found success rates below 50% even for leading agents and, in retail, less than a 25% chance of solving a task in all eight attempts.τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains · Shunyu Yao et al., arXiv:2406.12045, 2024
- An analysis of agent tests found too little attention to cost, inadequate separation of test tasks, and difficulty reproducing results; the researchers propose assessing cost and accuracy together.AI Agents That Matter · Sayash Kapoor et al., arXiv:2407.01502, 2024
What I think: I would measure accuracy, consistency, and cost before promising results, and repeat that evaluation as agents evolve.
10
Will AI agents replace employees?
AI agents can automate some tasks, but doing a task well is not the same as replacing an employee entirely.
According to METR, most jobs involve dealing with other people and producing results that cannot be scored automatically. Performance on well-defined tasks therefore does not, by itself, show that an agent can take on all the responsibilities of a job.
- METR warns that even an eight-hour task-completion horizon would not mean AI could automate all jobs, because its tests mainly cover well-specified technical work.Task-Completion Time Horizons of Frontier AI Models · METR, METR (last updated 8 September 2026, no longer maintained), 2026
- A study called the Remote Labor Index tested agents on real remote-work projects across several sectors, and the best agent automated only 2.5% of projects from start to finish.Remote Labor Index: Measuring AI Automation of Remote Work · Mantas Mazeika, Alice Gatti, Cristina Menghini et al., arXiv:2510.26787, 2025
- A study called GDPval found the best models approaching experts in the quality of their deliverables and analyzed their potential to perform those tasks faster and more cheaply with human oversight.GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks · Tejal Patwardhan, Rachel Dias, Elizabeth Proehl et al., arXiv:2510.04374, 2025
- A preliminary Stanford study found no widespread job losses, but employment among workers aged 22 to 25 in AI-exposed occupations was 19% below the trajectory of less-exposed peers, without treating this as proof of cause.Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence · Erik Brynjolfsson, Bharat Chandar, Ruyu Chen, Stanford Digital Economy Lab, 2025
What I think: I recommend evaluating agents on real work before promising replacements and pursuing gains that support employees without treating them as disposable.
Who writes this
I’m Vitor Araujo, an engineer. I build AI products for companies, and I speak and write about AI ethics and the philosophy of technology. More about me.
Other guides









