AI agents vs chatbots: the difference for your work

What a chatbot does

A chatbot works through exchanges. The user types a request, the model writes an answer, and nothing more happens until the next message. The chatbot remembers the conversation in progress.

Between two messages, the work falls to the user. The user finds the documents, pastes them into the window, chooses the next question and compares each answer with the sources.

Google Cloud separates three categories of tools: bots, AI assistants and AI agents. In its comparison table, the assistant is reactive. It responds to requests, can recommend actions and leaves the decision to the user. The agent is proactive and works towards a goal. It can carry out complex actions in several steps and make decisions on its own.

What an AI agent does

The major AI companies define the agent by its independence. OpenAI’s practical guide to building agents describes them as systems that independently accomplish tasks on the user’s behalf. Anthropic describes systems in which the model directs its own process and its own use of tools.

In practice, an agent receives a goal and breaks it into steps. At each step, it chooses a tool, such as a web search, the reading of a file or a calculation in a spreadsheet. It reads the result and decides on the next step. According to OpenAI, the agent recognises when the task is complete and corrects its actions when needed. When it fails, it stops and hands control back to the user.

OpenAI’s guide places simple chatbots and single-turn applications outside the category of agents, because the model does not direct the course of the work.

Time is the other difference. A chatbot answers in a few seconds. An agent can work for an hour on a file, or follow a subject for several weeks and report what changes.

Several agents on one task

A long task can be divided between several agents. Each agent has its own role. One agent searches for documents while another checks the figures and the quotes.

OpenAI’s guide describes two set-ups. In the first, a manager agent coordinates specialised agents. In the second, agents of equal rank pass tasks to one another according to their speciality.

OpenAI advises starting with a single agent. The company recommends adding agents when that single agent stops following complex instructions or keeps choosing the wrong tools. Dividing the work has another benefit. One agent can check the work of another before a result reaches the user.

Engineers already hand entire projects to agents

Software development is the field where agents are most advanced. According to GitHub’s documentation on 17 October 2026, its Copilot cloud agent researches a code repository, plans changes and implements them in the background. It edits files, runs tests, creates a branch and pushes its changes. The developer reviews the result and can ask for corrections before accepting the changes.

In February 2026, an Anthropic researcher gave 16 agents a single project. They had to write a C compiler in Rust from scratch, able to compile the Linux kernel. The agents worked in parallel on the same code base, without active human intervention, over nearly 2,000 sessions. The result is a compiler of about 100,000 lines that builds Linux 6.9 on three processor architectures. The researcher lists its limits. The code it generates is not very efficient, and it cannot yet replace an established compiler.

In March 2025, METR, a research organisation that measures AI abilities, published a trend. The length of the tasks that agents complete on their own in half of their attempts, measured by the time a professional needs for them, had doubled about every seven months over the previous six years. METR has since updated some of these estimates.

Engineers now delegate whole projects to agents and review the result. Verbatim brings the same way of working to research, analysis and writing.

When a chatbot is enough and when you need agents

Anthropic advises finding the simplest solution and adding complexity only when it is needed. According to its guide, agent systems accept more delay and more cost in exchange for better performance on the task. They suit open tasks whose steps change from one case to the next.

A chatbot is enough when:

• the question fits in one message
• you supply the text to work on
• the answer can be checked in a few minutes
• nothing has to happen once the conversation ends

Agents are the better choice when:

• sources have to be found before they can be used
• the material runs to hundreds or thousands of pages
• a subject has to be followed for days or weeks
• the same result needs several checks
• the work has to continue while you do something else

To decide, count how many times you copy a result from one tool to another between the question and the finished work. Beyond a handful of copies, the task has several steps, and an agent can carry them out one after another.

The risks of agents

Anthropic names the first risk in its guide to building agents. The autonomy of agents brings higher costs and the potential for compounding errors. An error made at the second step carries into every step after it. A date misread in a document can end up in a timeline and then in the final text.

The second risk concerns sources. An agent that has read 300 pages delivers a summary of three paragraphs. If the summary does not show the passages it relies on, a misreading stays hidden until someone opens the original. Language models can also produce statements that appear in no document. This is called an AI hallucination.

The third risk concerns actions. An agent that sends messages or publishes content acts in the user’s name. OpenAI recommends human oversight for sensitive, irreversible or high-stakes actions. GitHub states that the agent’s logs do not replace the developer’s own review and testing.

What to demand from an agent

Check six points before you hand a task to an agent.

• Every statement links to its source, with the exact passage.
• Every quote is copied from the original document, never from a summary.
• The agent waits for your approval before it sends, publishes or deletes anything.
• Every step is recorded, including what was searched, what was read and what was set aside.
• You can stop the work, correct a step and resume from that point.
• The provider states where your documents are stored and whether they train its models.

Anthropic describes agents that pause at checkpoints to get human feedback. For work based on facts, these checkpoints cover every fact and every quote.

AI agents with Verbatim

Verbatim gives you a team of AI agents, from the first lead to the final draft. You describe your subject in plain words. The agents search, read, follow your sources and write in your style, for hours or for weeks. Every fact stays tied to its source, and you make every call.

Several agents work in parallel on the same project. Each one has its own role, and they check one another’s work.

• The monitoring agent follows hundreds of sources day and night, including the press, public registers, official decisions, social networks and the sites you add. It flags what changes your subject, with the passage that proves it.
• The reading agent covers the web, public registers and up to 40,000 pages of documents. It answers with the exact page.
• The People agent finds who knows the subject and how to reach them, then contacts the sources.
• Before an interview, the preparation agent gathers what the person has already said, what the documents show and the questions they will try to avoid.
• Verification matches every claim with its source and the exact excerpt. Each claim is marked verified, contradicted, outdated or inconclusive. A quote found only in an AI summary never counts as verified.
• Writing follows your style, learned from your published texts, in every format from the article to the report.

You see what each agent is doing at every step. You can stop or undo anything, and nothing leaves the project without your approval.

Verbatim tests models from several labs on each task and keeps the one that does it best. Verbatim has built its own technology to tie every fact to its source. Its style engine was designed with linguists. Your projects are hosted and processed in Europe, and your content never trains a third-party AI. An iPhone app is on the way, along with Verbatim inside Claude and ChatGPT.

Journalism is the first field Verbatim serves, because it demands a source for every sentence. The same agents work for researchers, analysts and authors on their own material. VerbatimAgent, for one person, costs from €49.95 to €149.95 a month including VAT, with a 14-day free trial. VerbatimOS, for a newsroom or a team, costs from €109 to €249 a month, as set out on the pricing page.

Sources

• Anthropic, Building Effective AI Agents
• OpenAI, A practical guide to building agents (PDF)
• Google Cloud, What are AI agents? Definition, examples, and types
• GitHub Docs, GitHub Copilot on GitHub.com
• Anthropic, Building a C compiler with a team of parallel Claudes, 5 February 2026
• METR, Measuring AI Ability to Complete Long Software Tasks, 19 March 2025