Skip to main content
webforte

AI agents that work inside your systems

A chatbot answers questions, an AI agent completes a task: it plans the steps itself, looks up data in your systems and prepares the result. We develop AI agents for companies with fixed rules and human approval wherever a mistake would cost money.

  • Important actions are approved by a person
  • Every step the agent takes is recorded
  • The agent can only do what it is permitted to

AI agents at a glance

An AI agent is a program built on a language model that breaks a task into steps itself and carries it out with tools: it searches documents, reads and writes to a database, email, accounting system or CRM. A chatbot only answers, an agent acts and writes the result where it belongs. Classic automation needs input in a fixed form, an agent copes with free text. Webforte Technologies s.r.o. (company ID 23364343, Prague) develops custom AI agents for Czech companies, typically for enquiries, back-office documents and research. The agent has only the permissions it needs, passes checkpoints in code, a person approves important actions, and every step is logged. The price comes from a quote after the consultation.

What is an AI agent?

AI agent
An AI agent is a program built on a language model that receives a goal and decides for itself which steps will achieve it. It uses tools such as document search, a database or a company system, evaluates the results and continues until the task is done or a rule stops it.

Agentic AI is the general term for systems in which AI plans and carries out steps itself. When several agents work together under one controller (orchestration), it is a multi-agent system.

  1. Language model

    Understands instructions in everyday language and decides which step to take next.

  2. Tools

    Access to documents, databases and systems through an API, the interface programs use to talk to each other.

  3. Memory

    What the agent has already done, what it has found out and the context of the task.

  4. Rules (harness)

    Permissions, checkpoints, limits and human approval.

AI agentA model on its own only answers. It becomes an agent only with tools, memory and rules.

Chatbot, automation or AI agent?

All three are sold as AI, but they solve different tasks. Check which type your task really needs.

Swipe the table sideways to see all columns.

Chatbot, automation or AI agent?
Chatbot, automation or AI agent?ChatbotAutomationAI agent
What it doesAnswers questionsRuns a predefined sequence of stepsCompletes a multi-step task and chooses the approach itself
How it decidesFrom the question and the source materialBy rules someone wrote in advanceFrom the goal, the rules and the results of earlier steps
What input it handlesA question in everyday languageData in a fixed formatAn email, a document or an enquiry written in free text
ExampleTells a customer the delivery timeMoves every paid order into accountingReads an enquiry, looks up prices and stock and prepares a quote for approval
When it fitsRepeated questions from customers or employeesA stable process with input in a fixed formVariable input that needs judgement and work across several systems

An AI agent is not always the better choice. If classic automation can do the job reliably, it is cheaper to run and easier to check, and we will recommend it.

How AI agents work: orchestration and harness

A bigger task is not done by one universal agent but by several agents with narrow roles. A planner directs them (orchestration) and the whole thing is held together by a harness: software around the model that sets what agents may do, when they stop and who checks them.

Harness

  1. Input

    Brief

    An email enquiry, a document or a request from your system.

  2. Orchestration

    Planner

    Breaks the task into steps, assigns them to agents and tracks what is done.

  3. Specialised agents

    Each agent does one thing

    A narrow role is easier to test and check than one agent for everything. Steps that do not depend on each other run in parallel.

    • Research

      Finds the customer, order history and documents.

      Tools

      • CRM
      • document search
    • Calculation

      Works out the price from the price list and checks availability.

      Tools

      • price list
      • e-shop stock
    • Draft

      Writes the quote using your templates and tone.

      Tools

      • templates
      • email draft
  4. Check in code

    Checkpoint

    An ordinary program verifies the rules: the totals add up, the customer exists, the claim has a source and the agent has not exceeded a limit.

    • Passed: continues
    • Failed: back for fixing
  5. Human in the loop

    Approval

    Before the agent sends a quote or writes to the accounting system, a person reviews and confirms the result.

    • Approves
    • Returns with a note
  6. Output

    Result and log

    The output is written to your system. Every step, every tool used and every human decision stays in the log.

Rules around the agents

A language model does not police itself. Reliability depends on the harness we build around it.

  • Tools

    Precisely defined functions through which the agent works with your systems. If it is not among the tools, the agent cannot do it.

  • Permissions

    Set per tool, with reading kept separate from writing. Sensitive actions always go through approval.

  • Memory

    The state of the task and context from your data, so the agent does not repeat finished steps and picks up where it left off.

  • Record (log)

    Every step, tool and output. You can trace why the agent did what it did and fix the mistake in the rules.

  • Evaluation

    A set of test cases from your own work. We check every change to the model or rules against it before it goes live.

  • Limits

    A cap on the number of steps, time and cost per task. When the agent reaches it, it stops and hands the task to a person.

How to build an AI agent you can trust

Reliability does not come from the model alone but from the system around it. We use checks in code, sources behind claims and human approval in our own AI tools too.

Start with the task, not the technology

First we describe the task the way a person does it today: inputs, decisions, systems and the places where mistakes happen. Only then do we decide whether automation, one agent or several cooperating agents are needed.

  • Step-by-step description of the task
  • Decisions that must stay with a person
  • A clear measure of a good result

AI agent security and permissions

An agent that acts in company systems needs rules like a new employee, only stricter. We set this up for every agent we build.

Permissions
only what is needed, reading separate from writing
The agent gets only the tools and data the task requires. Preparing quotes, it may read the price list and create drafts, but not send emails or change prices. Writing is the first candidate for approval.
Sensitive actions
a person approves
Sending to a customer, posting to accounting, a payment or deleting: the agent prepares, a person decides.
Planted instruction
cannot bypass approval
Text in an email or document can contain a planted instruction that the model may not always recognise. So sensitive actions never depend on the model alone: they always wait for a person's approval.
Record
every step
The log holds the input, the tools used, the outputs and the person's decision. It serves for tracing errors and for evaluation.
Data sent to the model
minimum
We send the model only the data a given step needs. We choose the model and where data is processed according to the data the agent works with.

Process: from task to agent in operation

Every step has a deliverable that you can see and comment on.

  1. Consultation and choosing the task

    We discuss where you lose the most time and pick the task where an agent pays off first. We also check whether simpler automation would do.

    OutputRecommended solution and a non-binding quote

  2. Task map and harness design

    We describe inputs, steps, systems and decisions. We design agent roles, tools, permissions, checkpoints and the points where a person approves.

    OutputAgent design including permissions and checkpoints

  3. Test set

    From your real cases we build a set to measure the agent against, and define which mistake is unacceptable.

    OutputSet of test cases and success criteria

  4. Prototype on your data

    We build a first version and run it on the test set. We show you the results including the errors and adjust the rules.

    OutputEvaluation of the prototype on the test set

  5. Pilot with approvals

    The agent runs in live operation, but a person approves everything important. From the log we see where it goes wrong and refine the rules.

    OutputAgent in pilot operation with a record of steps

  6. Operation and improvement

    Based on the pilot we widen what the agent may do alone, while a person still approves sensitive actions. After every change we re-test on the test set and watch costs.

    OutputAgent in operation with regular evaluation

When an AI agent makes sense and when it does not

An agent makes sense where people retype and look up data across several systems and the input has no fixed form.

An agent pays off when

  • The task repeats, but the inputs have no fixed form: emails, documents, enquiries written in people's own words.
  • Someone currently retypes and looks up data across several systems.
  • Handling enquiries
  • Back office and documents

An agent does not make sense when

  • The process is the same every time and the inputs have a fixed format. Classic automation is enough: it is cheaper and easier to check.
  • Answering questions is enough. An AI chatbot built on your documents is better suited to that.
  • The task comes up only a few times a year and the development would not pay for itself.
  • A mistake would be irreversible and nobody has capacity to check the outputs.

Frequently asked questions

An AI agent is a program on a language model that receives a goal, plans its steps itself and uses tools: it searches documents, reads and writes to systems or prepares texts. Agentic AI is the term for systems in which AI plans and carries out steps itself. Every agent we build has defined permissions and checkpoints, and a person approves important actions.

A chatbot answers, an AI agent acts: it receives a task, breaks it into steps and carries them out in your systems. Classic automation runs a fixed sequence and works as long as the input matches the format, while an agent copes with free text too. Automation is cheaper and more predictable, so we recommend it wherever it is enough. We use an agent where judgement is needed.

Technically a language model, a set of tools and a loop in which the model picks the next step are enough. For business use, everything around it matters more: a task description, permissions, checkpoints in code, human approval, a step log and test cases from real practice. We go from a task map through a prototype on your data and a pilot with approval to live operation.

As safe as the rules you build around them are strict. The agent gets only the permissions it needs, a person approves sensitive actions (sending to customers, posting to accounting, payments), code checks the rules, every step is logged and a task has a limit on steps, time and cost. A planted instruction in an email therefore cannot trigger a sensitive action without approval.

The price depends mainly on how many systems the agent uses, how many agents and steps the task needs, how strict the checks are and how many tasks the agent processes, since model operation is paid by usage. We do not publish a price list, as agents differ case by case. After the consultation we prepare a non-binding quote for development and an estimate of running costs.

Usually not. The agent connects to what you already use through an API or through MCP, an open standard for connecting AI to data and tools. If a system has no interface, we look for another safe route, such as file exchange, or recommend a change. In the consultation we go through what you work with and tell you what can be connected.

Let's talk through your case.

Describe a task that someone currently does by hand across several systems. In a consultation we assess whether automation is enough or an agent is needed, where a person has to decide, and prepare a non-binding quote. We communicate in English.

Free consultation