AI agent development company whose agents get tested before they go live

AI agent development for multi-step work in your own systems. We write down the bar an agent must clear before we build it, then release in stages.

Illustration of an AI agent loop: a model planning steps, calling tools and databases, and pausing for a person to approveIllustration of an AI agent loop: a model planning steps, calling tools and databases, and pausing for a person to approve

AI AGENTS

What is an AI agent?

An AI agent is software in which an LLM works toward a goal. It plans steps, calls tools such as APIs, databases and internal systems, checks the result and repeats until the task is done or a person takes over. A chatbot answers a prompt. A script follows fixed steps. An agent chooses its next one.

Most of the engineering in an agent sits outside the model: tool use, orchestration between agents, MCP connections to your systems, and the point where a person approves what can't be undone. Without the acting part, an assistant that only answers from your documents is a smaller generative AI build.

Agents are one part of our AI development practice, next to generative AI, machine learning and AI integration. We start a build with discovery, and discovery is allowed to conclude that you don't need an agent. Plenty of agent projects should stop at that point.

Illustration comparing a chatbot answering once, a fixed script following set steps, and an agent choosing its next stepIllustration comparing a chatbot answering once, a fixed script following set steps, and an agent choosing its next step

WHEN NOT TO BUILD

When is an AI agent the wrong answer?

Fixed automation already does the job

If the steps are known and never call for judgment, a script or a rules engine does the job for less and gives the same answer every time. An agent would add a model's cost and its variance, and buy you nothing for either. Say your invoices always arrive in one layout from the same handful of suppliers. Write the script.

A RAG assistant would be enough

Where people only need answers from your documents and nothing in your systems has to change, an assistant that retrieves and answers costs less to build and to test. That is generative AI work. An agent earns its cost when something has to happen next: a record updated, a ticket opened, a refund issued.

The data isn't ready

An agent acts on what its tools return. If customer records are duplicated, stale or spread across three systems that nobody owns, the agent acts on bad data at machine speed. Then the first project is the data work, which our data practice takes on, and the agent comes after.

A mistake costs too much without a person

Some actions are hard to take back, such as paying a supplier, changing a contract or writing to a customer. If nobody can review them first, an agent is the wrong tool until an approval step keeps a person in charge.

Illustration of a decision fork: one path to a simple script, one to an assistant, one to an agent with a person approvingIllustration of a decision fork: one path to a simple script, one to an assistant, one to an agent with a person approving

WHAT WE BUILD

Agents we build, by kind of task

We sort agent work by the kind of task, from one agent with a few tools to several agents sharing a plan. The agent makes the judgment calls, the fixed steps around it stay ordinary workflow automation, and we set a limit on what each agent may do before a person has to approve.

Would you rather build agents in house? We staff AI engineers who build agents inside your team, vetted on RAG, evaluation sets, latency, cost and model drift. Across the company, our engineers average 7 years of experience, and 80%+ are senior-level or above.

Single-task tool-using agents

A short list of tools and one goal, such as looking up an order, checking the refund policy and drafting a reply for a person to send.

Multi-step workflow agents

The steps change from case to case, so the agent reads each request, picks the next action, calls the systems involved and logs what it did.

For example, sorting incoming support tickets by urgency and opening the matching customer record before anyone on your team replies.

Multi-agent orchestration

Several agents with separate roles, one planning and others researching or checking its work, coordinated in a framework such as LangGraph.

Each extra agent adds model calls, response time and failure points, so go multi-agent only when one agent with a checker step can't do the job.

Agentic RAG over your data

Retrieval inside the plan-and-act loop. The agent decides what to look up, goes back to your documents or databases when the first result falls short, and cites its sources.

Agents over internal systems

Wiring agents into existing systems means tool definitions, authentication and error handling for the CRM, ERP or finance system an agent reads and updates, through MCP or the system's own APIs.

For example, matching invoices against orders and passing the exceptions to a person.

GO-LIVE BAR

How do you know an agent is ready to go live?

A test set and a bar, agreed before the build

We write the bar down before we write the agent. With you, we collect real tasks whose good outcome is known and agree what the agent must reach on them. A useful bar names a pass rate, a cost per task and a response time, because an agent that is right but slow or expensive still fails the job. Run each case more than once, since paths vary. Later the same test set can compare models on those numbers, so a model swap is a measured decision. If the agent misses the bar, it doesn't go live.

Tool permissions and approval gates

An agent gets the access you would give a contractor on their first day, only the tools its job needs. Payments, messages to customers and other actions that are hard to undo wait for a person to approve them. Put a cap on steps and spend per task so a runaway loop stops, and have a tool call that keeps failing hand the task to a person. An email or document the agent reads may carry instructions meant to redirect it (prompt injection). Limited tools and the approval step limit the damage.

Tracing runs and scoring them

A trace records what the agent planned, which tools it called and what came back, so a failure can be followed to the step that caused it. Our engineers work with Langfuse, LangSmith and Ragas to trace runs and score them against the test set.

Monitoring after launch

Live inputs drift away from the test set. Customers word requests in new ways, and a new product line brings questions the test set never covered. Watching a live agent for drift and checking its quality after launch is work our data scientists take on.

Illustration of an agent's run traced step by step against a checklist, with a gate a person opens before releaseIllustration of an agent's run traced step by step against a checklist, with a gate a person opens before release

PROCESS

From discovery to a released agent

Discovery is run by an architect, a business analyst and a delivery manager, plus a lead engineer if needed. Your side brings the people who do the task today and access to the systems the agent will touch. Arrange that access early, because an agent can't be tested on systems it can't reach.

  1. STEP 1

    Discovery phase

    We map the task, your systems and your data. Where you can share data under an NDA, we run the riskiest part on a sample of it before you commit to a build.

  2. STEP 2

    Agree the test set and go-live bar

    The people who know the work today say what a good outcome looks like, and we turn that into test cases and a written bar.

  3. STEP 3

    Build in Agile sprints

    We build in sprints with bi-weekly demos, tracked in Jira and Slack, and run the agent against the test set, so progress shows up as results.

  4. STEP 4

    Staged release with a person approving

    The agent goes into your systems in stages. At first a person approves its uncertain or hard-to-undo actions, and it gets more autonomy only as its results allow.

  5. STEP 5

    Support after launch

    Fixes, updates and small improvements by priority, with response times agreed in your support contract.

Which task would you hand to an agent first?

Describe the task, the systems it touches and who checks the result today. Whether it needs an agent is the first question we look at.

TECHNOLOGIES

Tools our engineers work with on AI agents

OpenAI API

Anthropic Claude

Google Gemini

Mistral

Amazon Bedrock

Azure OpenAI

Google Vertex AI

LangGraph

MCP

Langfuse

LangSmith

Ragas

Python

HOW WE WORK

Engagement models and terms

Project delivery

We deliver an agreed scope end to end: plan, team, milestones, QA and release. You make the product decisions and accept each milestone.

Fixed price per milestone after discovery, or time and materials against an estimate.

  • Scope changes go through change requests.
  • On fixed-price projects, a warranty period agreed in the contract, usually 30–60 days.

Discovery phase

A short phase before a build: requirements and backlog, architecture, an estimate and a delivery plan.

Fixed price, usually 2–4 weeks.

  • The outputs stay with you whether or not the build continues with us.

Staff augmentation

Our engineers join your team and work in your process and tools. You manage them; we source, vet and contract them.

Hourly rate per engineer by role and seniority, invoiced monthly.

  • If an engineer isn't the right fit, we replace them within 14 days at no extra cost.
  • Knowledge transfer to the replacement engineer isn't billed.
  • There is no free trial period.
  • 30 days' notice to take an engineer off or end the engagement, with no minimum term beyond it.
  • Out-of-hours on-call only if agreed with you; we set its hours and billing together.
  • After two years, moving an engineer onto your own staff is something we can discuss.

Whichever model you choose

  • We reply to an inquiry from the contact form within 24 hours.
  • We sign an NDA before work starts, and IP passes to you on payment.

Who signs off on what the agent does?

Tell us which actions the agent would take and what a wrong one would cost. That decides where the approval steps go.

FAQ

Questions buyers ask before they commission an agent

What does an AI agent cost to build and to run?

The build cost moves with the scope of the task, how ready your data is, how many tools and systems the agent touches and how much testing the bar calls for. Running cost comes on top: model usage for every task the agent handles, which is why a good bar names a cost per task. As part of the build we set up per-call cost tracking and budget enforcement, with limits and alerts, so the model bill can be read task by task.

How long does it take to get an agent live?

Discovery usually takes 2–4 weeks and ends with an estimate and a delivery plan, which is when the build time for your scope gets set. Typical ranges for early builds are up to 1 month for a proof of concept, up to 2 months for a proof of value and 2–4 months for an MVP, and none of them is a commitment. A proof of concept or a proof of value can run without discovery first.

What access to our systems and data do your engineers need?

Named accounts that you issue in your own identity and access management, each limited to what the engineer's role needs and removed when they leave the engagement. Where we handle your data, a data processing agreement is part of the engagement. Systems we control are hosted in the EU, while code and data in your own repositories and cloud stay wherever you keep them. We work under GDPR as an EU company. An AI build can call its models through your own Amazon Bedrock, Azure OpenAI or Google Vertex AI account, so the model runs under your account in your cloud.

Can you take over an agent another team started?

Yes. It starts with an audit: we read the code, its dependencies and how it is deployed and released, and look for missing tests and documentation. You get a written list of findings with what we would fix first. A short handover follows as its own phase, covering access, environments, runbooks and a walkthrough from the previous team where one is available. If the agent came with no test set, that is the first gap to close, because without one nobody can tell whether a fix made it better or worse.

Can you show us an agent you have built?

No agent case study is published yet. Our AI and data-science work to date sits under client NDAs, which is why only two AI case studies are published. The nearer one to agent work is a generative AI system that drafts an R&D tax consulting firm's reports on Azure OpenAI, in the firm's own Azure account. It drafts each report for a consultant to edit and sign off, so it is generative AI work, not an agent. Its case study is linked from our generative AI page. The other, demand forecasting for a logistics company, is predictive machine learning. What we can show for your agent is how we would set the test set and the bar for your task, and how the controls described on this page would apply to it. Whoever you talk to, ask how they would score the agent before it goes live and what happens when it misses.

Does the EU AI Act apply to an agent you build for us?

It can, even if your company is outside the EU. The EU AI Act, Regulation (EU) 2024/1689, reaches providers and deployers outside the EU when an AI system is placed on the EU market or its output is used in the EU. It sets obligations by risk level, with stricter ones for higher-risk uses. Whether your agent falls under it, and at which level, is a question for your legal team to confirm.