AI agents · Assistants · Task automation

AI agent development: the agent acts, people set the limits

In AI agent development the main question is authority: what the agent is allowed to do, how its work is checked and what happens when it makes a mistake. An agent differs from a chatbot in that it acts: it chooses the steps, calls your systems and carries the task through to a result.

For teams that need to take repetitive work off people when that work still requires understanding the content: parsing requests and documents, first-line support, preparing reports, routine in CRM and internal services.

What is included

Scenario and limits of authority

We define what the agent does on its own, what it proposes to a person and what it never does. Its permissions are set in the architecture, not as a request in the prompt.

Tools and access to systems

We connect the agent to your databases, APIs and services and handle failures and timeouts in advance. Sooner or later a tool will return an error, and at that moment the agent must behave predictably.

Company knowledge

Documents, internal procedures and data become the agent’s memory through semantic search and knowledge graphs. The agent then works from your rules and does not have to guess.

Quality evaluation

Test task sets and metrics are in place before the rollout to production. Without them, “it got better” stays a feeling, and every prompt change becomes a lottery.

Audit and observability

Every agent action is logged: what it did, on what grounds and what it cost. The log serves both incident investigation and control of spend on model calls.

Interfaces

The agent works where people already work: a messenger, a web interface, an embedded widget, an internal system. A separate service that people have to make a point of visiting loses its users.

How we work

  1. Choosing the task

    We pick a scenario that repeats often and has a verifiable result. Rare and one-off tasks do not go to an agent.

  2. Prototype with evaluation

    We build the agent together with a set of test tasks: quality is measured from day one, before the first user complaints.

  3. Pilot with people

    We launch on a real flow of work, and a person confirms every action the agent takes. We look at where the agent makes mistakes and on what.

  4. Production

    We give the agent more independence where quality is confirmed and keep human confirmation where a mistake is expensive.

  5. Operation

    We monitor quality and cost, review new cases and rerun regression checks when the model changes.

Why us

AI agents in our own work

We do our engineering work together with AI agents from several vendors, under one set of rules. People start the sessions and accept the results, so we know well where an agent goes wrong and how to catch it.

An agent as ordinary software

Versions, test task sets, metrics, regression checks when the model or the prompt changes. We maintain an agent the way we maintain any program that has to run in production.

Working on-premise

If your data cannot leave the perimeter, we deploy the agent on local open-weight models inside your infrastructure.

Case studies

Aviation · aircraft components trading and MRO

Pilot

Aircraft parts procurement: from email to a managed pipeline

An AI agent reads incoming RFQs, compares supplier quotes and checks trace documents. A separate deal-tracking system moves each deal through mandatory stages.

Full loop
the AI loop ran end to end on a live procurement process, from an incoming RFQ to a team notification. This was a pilot and the figures are our own so far: the client has not confirmed them yet
3
types of trace documents checked by the agent: FAA 8130-3, EASA Form 1 and COC. Like the figure above, it still awaits the client’s confirmation

Distribution · Partner networks

In production

A sales and support platform for a distributed partner network

Partners ask who on their team is inactive and what blocks the next level, and get numbers for their own part of the network only. Access is enforced at the SQL level.

Their own slice
a partner gets figures for their own team only: a filter on the current user is added to every SQL query
6 layers
of data protection between the language model and the database, from blocked personal-data columns in SQL to rate limits; counted in the code

AI infrastructure · multi-agent systems

In-house product

Agent teams assembled from configuration, with paired critic agents

The core of a multi-agent platform in 2.5 weeks: 29 declarative agents in three teams, agent-critic pairs, shared memory on four databases, five autonomy levels.

25,000+
lines of Python in the platform core across 134 modules, counted from the repository
29
declarative agents in three ready-made teams (engineering, cognitive, research); counted from YAML specs

IT consulting · Lead qualification

In-house product

A pilot plan for the visitor, a qualified lead for sales

We replaced the contact form with an AI consultant: it interviews the visitor, returns a pilot plan with KPIs and risks, and sales gets a lead with budget and timeline.

1 day
to build a working MVP: frontend, backend and deploy configs
16 days
from first commit to production with HTTPS and an issued certificate on its own domain

Frequently asked questions

How is an AI agent different from a chatbot?

A chatbot answers with text, following a preset script or using a model. An agent performs a task: it plans the steps, calls external systems, checks the result and sees the job through. For an agent, conversation is only the interface; the work itself happens in your systems.

How much does it cost to develop an AI agent?

Most of the cost goes into integrations, preparing the knowledge and checking quality; the agent itself costs less. A simple agent with one data source and one integration point costs noticeably less than an agent with access to several systems that has to run on-premise. We give a price range after reviewing the scenario.

What happens if the agent makes a mistake?

Whatever the architecture provides for. Actions with costly consequences need a person to confirm them; the rest are logged and reversible. We design the system on the assumption that a mistake will happen, because it will.

Can the agent run on a local model?

Yes, if your environment requires it. Local models are weaker at complex reasoning, so we match the scenario to what the model can do and discuss this up front.

How do you measure the quality of an agent’s work?

With a set of test tasks whose correct answer is known, plus metrics on the real flow: the share of tasks completed without intervention, the share corrected by a person, the cost of one task. The same metrics show when the agent can be given more independence.

Do we need a multi-agent setup straight away?

Usually not. One agent with good tools is cheaper and more predictable. Several agents make sense when you need different roles with different permissions, or when one participant checks another’s work. The page on multi-agent systems covers this in detail.

Work it out on your own figures

The AVA calculator works out the economics of your process from your own figures, with no form and no price from us. If automation does not pay off, it says so.

Estimate the payback

Tell us what you are dealing with

Describe the problem in your own words. If it can’t be solved or won’t pay off, we’ll say so before any work starts.

In the reply we will say what we understood, what we need to clarify and whether it is worth doing.

or email us directly: hello@xteam.pro

See case studiesNew write-ups on Telegram