Agents that hold up in production

A language model that fetches the right information, drives your tools, then hands control back at the right moment. Not a demo that impresses once and breaks the next day.

  • From €5,000
  • 6 to 10 weeks
  • Remote, Europe
Agents shipped for
Limova, AI scale-up
Kiro, AI scale-up
Kepler Cheuvreux, finance
OECD, international institution
Limova, AI scale-up
Kiro, AI scale-up
Kepler Cheuvreux, finance
OECD, international institution

What you
get.

Included in the service

  • Identifying the tasks where an agent beats a plain form
  • Document search, so the agent answers from your own material instead of making things up
  • Connecting your tools: sending an email, reading an order, writing into your software
  • Guardrails: what the agent can do on its own, what always goes back through a person
  • Quality checks kept up to date, so you see the agent drift before your users do
  • Cost tracking by usage, so the bill never comes as a surprise at month end
  • Documentation, a walkthrough video, then a handover session

Not included

  • Training a language model from scratch
  • Writing your content and sales copy
  • Branding and visual identity work
  • Connecting to your internal tools when access is not ready
  • Ongoing support and changes once the project is delivered, handled separately

How
it works.

  1. 01Week 1

    We isolate the task worth automating

    An agent only earns its keep on a precise, repeated task where the answer depends on context. We name it before writing a line: what it must read, decide, then do.

  2. 02Weeks 2 to 4

    The agent reads your documents, not its own

    Document search connects the agent to your texts, your records, your history. It answers from what actually exists in your company, and says so when the answer is not there.

  3. 03Weeks 5 to 8

    The agent acts, under watch

    It triggers your tools one at a time, with a guardrail at every sensitive step. A quality check runs continuously to catch drift before a customer sees it.

  4. 04Weeks 9 to 10

    You get everything back

    The code, the access, costs measured by usage, a walkthrough video, then a session with whoever takes over next.

years_freelance
+10
Years freelancing, remote for European teams.
missions_shipped
+170
Missions delivered, each shipped to production.
companies
+31
Startups, SMBs, enterprises.

What users say.

Yet another successful mission with Pierre.
a
Alexandra
Founder · Digital Theory
Serious, committed, methodical, with a strong technical background. He's someone who will deliver your projects with real professionalism. Highly recommended.
a
Alexandre Gorski
Freelance developer

The questions
I always get.

How much does a production agent cost?

From €5,000. The amount depends on how many tools are connected and how much guardrail work is required. A firm price lands after the first week, once the task is isolated.

  • The price is set before development starts, never adjusted along the way.
  • Connecting one more tool or adding a guardrail is priced separately, on request.
  • The first week of analysis is billed even if you stop there.
What does document search actually do?

It lets the agent answer from your own material instead of guessing. Before answering, it looks up the relevant passages in your documents, then builds its answer on top of them.

  • The agent cites what it read, so you can check the source.
  • If the answer is not in your documents, it says so instead of making one up.
  • Adding a new document updates what the agent knows, no retraining needed.
What does tool calling actually do?

It lets the agent act instead of only replying. Instead of telling you what to do, it sends the email, reads the order, or writes into your software on your behalf.

  • Every connected tool carries a rule: the agent acts alone, or always goes back through a person.
  • A sensitive action always passes through a validation step before it runs.
  • The full list of what is allowed is set out above on this page.
How do you know when the agent is drifting?

Quality checks run continuously on real cases, not just at launch. A drop shows up before your users notice it.

  • Checks focus on cases that already went wrong, so they don't happen twice.
  • A dashboard shows the trend week by week.
  • An alert fires when a threshold is crossed.
How do costs stay under control?

Every call to the language model has a cost, and it climbs fast without watching. Usage tracking shows who costs what, and lets you cut a drift before the invoice arrives.

  • A cap per user or per task avoids a bad surprise.
  • Cost detail lives in the dashboard delivered with the project.
  • A cheaper model replaces the strongest one wherever the difference is not visible.
Do you work remotely?

Yes, from Paris, for clients across Europe. I work in French, English and Italian. Weekly calls adapt to your timezone.

Thirty minutes to know
if an agent is worth it here.

Tell me what task you want to hand to an agent. I will tell you if it holds up, what it needs in guardrails, then what it costs.

Book 30 minutes