Atos put 400 engineers in a three-day agentic AI league

You've probably sat through an agentic AI briefing where nobody in the room has ever configured a guardrail. The slides are confident. The hands are not. Atos and Amazon Web Services ran the opposite: a scored league. You built the thing. A leaderboard punished you for being slow and wasteful.

Four hundred engineers. Three days. Live services. Not a course. Not a recording.

The room was not full of experts

The incoming mix, from the blog, was that 5% had no prior knowledge of agentic AI, 25% had basic awareness, 50% understood it and had never put a hand on it, and 20% had practical experience. That is a real enterprise room. If your workshop assumes the 20%, you have designed a meetup.

The time they actually asked for was a two-hour kick-off, an hour of office hours each day and a one-hour finale. The rest happened around the day job. I like that more than a five-day offsite that the delivery team cannot attend.

The stack they used was Amazon Bedrock, Bedrock AgentCore, Lambda, Guardrails, Kiro and SageMaker. The challenges were not "write a poem." Guardrails. Memory. Pathfinding. Code execution. Retrieval. Fine-tunes. You can argue with the dungeon-maze metaphor. You cannot argue with "your over-eager filter just blocked the legitimate query."

Names, because this is not a nameless enablement blob

The top three were James Ponter, Adam Różewicki and Eduard-Cosmin Socol. Ponter won. He is Head of Hyperscalers for UKI Cloud and Infrastructure at Atos. His quote, which I will keep because it is the brief, is that academic learning gives you the foundation and the league puts it under pressure in a way that changes how you think. You cannot look up the answer when the clock is running.

Chris Byrne, Global Head of AWS Alliance at Atos, put it in a sequence: DeepRacer for reinforcement learning, AI League for fine-tunes, this year agentic AI. Practice without the live-client gun to your head. Atos also talks about Sovereign Agentic AI Studios in multiple locations. The public account names that sequence. It does not list a street address for each studio.

Four hundred engineers with registration data gained hands-on experience. There is no utilisation rate beside that number.

If you are buying AI training in the United Kingdom, this is the shape that transfers: a scored build, real services, mixed seniority and hours that fit around delivery. A two-day sprint that leaves a decision is closer to this than a vendor's awareness track.

The incoming mix is the brief for whoever designs your next academy. Half the room already "got it" and had never built anything. That is not a failure of hiring. That is what "we sent them a course" produces. A league with a clock is how you find out who can ship. Ponter's line about not looking up the answer when the clock is running is the opposite of most corporate LMS design, which is built so everyone passes.

Keep the scheduled hours short on purpose. Two hours plus office hours plus a finale is something a delivery team can attend. Five days offsite is something they will skip. Byrne's sequence (DeepRacer, then fine-tunes, then agents) is a curriculum even if they never call it that.

Ponter is a UKI hyperscalers lead. The winner is not a graduate hire who had a free week. That is the signal for anyone buying training in London: the people who own the alliance showed up and got scored. If your academy cannot get those owners into the room, you will train the people with spare time.

What this means for L&D

Score the room. Publish the skill mix. Cap the scheduled hours. Score efficiency, not just whether it worked. Name the winners internally so the rest of the company has someone to call. That is training that can survive a client call.

Sources: AWS Machine Learning blog, 1 September 2026.