Pascal EK Loui · AI Product Builder · Claude Expert · 16 agents in production · 100+ Claude skills · Claude Agent SDK · MCP · RAG · Fixed price from €3,500 excl. VAT · Paris and remote.
An AI agent orchestrator breaks a business process down into tasks, assigns each task to a specialized AI agent, connects the agents to one another and places a human checkpoint wherever a mistake is costly. They write the briefs, define the success criteria, check the outputs and correct them. The model executes. The orchestrator decides.
Depending on who you ask, the term is either intimidating or faintly amusing. In practice, it is the role I play every day on my own platforms and on client engagements. In 2026, the dividing line is no longer between people who use AI and people who don't. It runs between those who orchestrate agents and those who ask a chatbot questions.
Proof, not promises. 16 Claude agents in production. More than 100 reusable skills. Three platforms built solo on this approach: sceniq.studio, five agents that turn a brief into a product reel in 48 hours. tondroit.fr, a RAG pipeline that answers questions across eleven areas of law, citing the relevant legislation. DevisGo, a WhatsApp agent that generates a quote from a voice message.
Four moves, always in the same order.
Configuring the workflow is a small part of the job. The rest is briefing, checking and correcting. An agent is a very fast collaborator that needs a pilot. It is not a system you just leave running.
The titles sound alike. The jobs don't.
| Role | Works on | Delivers | Limitation |
|---|---|---|---|
| Prompt engineer | An instruction, a conversation | Better answers | Nothing runs without them |
| AI developer | Model code and infrastructure | A technical building block | Doesn't scope the business process |
| No-code Product Builder | Assembling tools (Make, Airtable, Webflow) | A working tool | Stops where the tool stops |
| AI agent orchestrator | The whole process, broken down into tasks | An agent workflow that runs under human oversight | Only as good as the scoping |
I combine the last two roles with product scoping. That is what I call an AI Product Builder. Agent orchestration is the execution side of it.
The question is no longer whether you need agents. It is why most pilots never reach production. The causes are rarely technical. Vague scope, unprepared data, underestimated integrations, no human checkpoint. A poorly scoped agent gets things wrong quickly and confidently. Preventing that is exactly the orchestrator's job.
| The task | What the agent does | What I keep |
|---|---|---|
| Interview research and synthesis | Transcribes, codes the verbatims, clusters the insights, produces the synthesis. | The fieldwork, and choosing the insight that changes the product. |
| Specs and user stories | Writes the specs, acceptance criteria and edge cases. | Journey consistency and priority trade-offs. |
| Working product | Codes the interface and the logic, connects real data. | Testing with real users within 72 hours, and the go/no-go call. |
| Content and videos | Five agents turn a brief into a product reel in 48 hours. | Art direction and messaging. |
| Sourced customer answers | A RAG system answers, cites the law and can admit when it doesn't know. | Choosing the sources, the fallbacks, the accountability. |
| Quotes and business documents | A WhatsApp agent generates the quote from a voice message. | Final sign-off by the tradesperson. |
I review your workflows, identify what can be automated, and estimate the gain and the risk. You leave with a prioritized automation plan, ready to build, and a list of what must stay human.
A working agent, tested on your real data, with real users. Scoping the hypothesis, a first version on your data, testing, and a costed go/no-go recommendation. You know whether it's worth deploying before you pay for it.
I join your team as an AI Product Builder. Scoping, design, specs, code, agents, deployment, monitoring, handover. One person accountable all the way to production, with no handoffs between four different disciplines.
A short program for designers, product managers and business teams. Qualiopi-certified training provider, eligible for OPCO funding. Every participant leaves with an agent that runs.
Most AI agent orchestration offers stop at a chatbot POC. I have agents in production and entire platforms built solo on this approach. One point of contact between the business process and the code that runs. Fifteen years of product judgment before AI, on demanding accounts: Chanel, Dior, Accor, Carrefour, Forvis Mazars, Microsoft. That is how you recognize a Claude expert. Shipped products, failures, a career that predates the models.
An AI agent orchestrator breaks a business process down into tasks, assigns each task to a specialized AI agent, connects the agents to one another and places a human checkpoint wherever a mistake is costly. They write the briefs, define the success criteria, check the outputs and correct them. The rare skill isn't launching an agent. It's knowing when to stop it.
A prompt engineer optimizes an instruction to get a good answer from a model. An AI agent orchestrator designs a system: several agents, tools connected through MCP, data, stop rules and human checkpoints. The first works on a conversation. The second delivers a process that runs without them, under their supervision.
In my case, the Anthropic ecosystem: the Claude API and the Claude Agent SDK for the agents, the Model Context Protocol (MCP) to connect tools and data, reusable skills to package know-how, RAG for sourced answers and Claude Code to ship the code. Around that, Next.js, Supabase and Vercel for production.
A one-week automation audit, priced on quote. A 72-hour AI Sprint from €3,500 excl. VAT, with a working agent tested on your data. A 6-month to 1-year embedded engagement at a fixed price for a complete product. Never billed by time spent: the budget is approved before work starts.
Rarely because of the model. Vague scope, unprepared data, underestimated integrations, no human checkpoint. A poorly scoped agent gets things wrong quickly and confidently. That is why orchestration starts with scoping and success criteria, not with choosing a model.
As few as possible. One agent per clearly defined task, not one agent per idea. On sceniq.studio, five agents are enough to turn a brief into a product reel in 48 hours. On tondroit.fr, a RAG pipeline covers eleven areas of law. More agents means more points of failure and a higher verification cost.
Based in Paris, on site for the kick-off and scoping workshops, remote for production. Clients in France, the United Kingdom, the United States and Asia, with an active company in Hong Kong. Reply within 48 hours.
45 minutes to scope the need, tell you what can and can't be automated, and put a price on the next steps. The first call is free. Reply within 48 hours.
Or by email: e.loui@uxdesignparis.fr · Phone: +33 6 35 31 50 35