AI agent development

AI agent development, with guardrails and human sign-off

An AI agent is software that uses a language model to decide on a sequence of steps and then carries them out through tools: it reads a ticket, looks up the order, drafts the refund and asks a person to approve it. The model is the easy part. The real work is deciding which actions the agent may take, what it must ask before taking them, and how you find out when it gets something wrong.

Book a scoping callSend us a project brief

Operated by Azinove SAS (France); we work with US teams remotely.

We build agents for teams that want that work done properly: actions limited to what the job needs, approvals wherever money, customers or records are at stake, and logs that let you replay any decision the agent made.

Agents we build

We start with one workflow that has a clear owner and an output you can check, take it to production, and only then widen the agent’s scope.

  • Agents that work in your existing tools through their APIs: CRM, help desk, ticketing, internal databases
  • Research and drafting agents that gather sources and prepare work for a person to finish
  • Back-office agents that route, classify and fill in records, with a review queue for unclear cases
  • Agent access to your documentation and knowledge through MCP servers or retrieval

Guardrails are the design, not an add-on

Every tool an agent can call gets an explicit permission, input validation and a limit. Actions that change records, send messages or spend money go through a human approval step until test results justify relaxing it. Each run leaves a full trace: the instructions, the tool calls, their results and who approved what.

Evaluation before an agent touches production

Before launch we build a test set from real cases, the awkward ones included, and measure the agent against it. The same set runs again whenever the instructions, the model or a tool changes, so a regression shows up in testing instead of in front of a customer.

What we have built ourselves

GitDoc, our documentation platform, ships a built-in MCP server that lets Claude Code and other agents read and write documentation pages. Building it showed us how agents behave with real tools: which calls they misuse, which permissions they actually need, and how to make their changes reviewable. Botsmith, our managed website chatbot, is in development and applies the same handoff rules to customer conversations.

Running the agent after launch

Agents need looking after: model updates, new edge cases, cost tracking and the occasional rollback. We can operate the agent with you, or hand over runbooks and the evaluation suite so your own team can.

Questions

What clients ask about AI agent development.

Another question? Email hello@cleverai.ai

What is the difference between an AI agent and a chatbot?

A chatbot answers. An agent also acts: it calls tools, changes records or starts workflows. That is why agents need permissions, approvals and audit trails that a question-answering bot does not.

Which agent frameworks do you use?

We pick per project. A small amount of our own orchestration code over a model provider’s tool-calling API is often easier to test and maintain than a large framework. We explain the trade-off for your case before we start.

Can an agent work with our internal systems?

Yes, if they have an API or a database we can reach securely. The agent connects through a service account with the narrowest permissions the job needs.

How do you stop an agent from doing something harmful?

Scoped tool permissions, validated inputs, limits on volume and spend, and human approval for any action that is hard to undo. Every run is logged, so you can see what happened and why.

Do we need a lot of data to build an agent?

Not for training, since agents run on existing models. You do need real examples of the task, to write the instructions and to build the test set that tells us whether the agent is good enough.

Can we start with a single workflow?

That is what we recommend. One workflow with a clear owner and a checkable output, taken all the way to production, teaches more than a broad agent platform built in advance.

Who owns the agent we build?

You do: the code, the instructions, the evaluation sets and the documentation are handed over at the end of the engagement.

Related

Services, products and reading related to AI agent development.

Talk to the team

Tell us about the AI agent development project you have in mind.

Operated by Azinove SAS (France); we work with US teams remotely.

Book a scoping callSend a project brief