From theory to delivery: How Atos upskilled 400 engineers in agentic AI
Amazon Web Services Rajesh Babu Nuvvula
Atos put 400 engineers through a three-day AWS agentic AI contest. The point wasn’t theory — it was building working systems under pressure, with scores for speed and efficiency.
Based on reporting by Amazon Web Services, Rajesh Babu Nuvvula — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Atos wanted its engineers to do more than talk about agentic AI. So, with AWS, it ran a three-day AI League for 400 people and pushed them into building real systems instead of sitting through another neat slide deck.
The mix inside the room was wide. Some participants already knew AWS. Others were using it for the first time, and some came from less technical roles like product owners and project managers. A small slice had never touched agentic AI at all, while most had either only basic familiarity or theory without hands-on work.
The challenge was a dungeon maze. Engineers had to build an autonomous agent that could find a route, solve problems on different tiles, dodge traps, collect coins, keep its lives, and reach treasure before time ran out. The scoring system rewarded not just success, but restraint: fewer wasted tokens, fewer unnecessary tool calls, and better efficiency all mattered. Fine-tuned small models also earned bonus points.
That design forced real trade-offs. Teams had to think about pathfinding, guardrails, memory, code execution, web scraping, and how many agents to split a system into. The AWS stack behind it was broad: Amazon Bedrock, AgentCore, Lambda, Kiro, SageMaker, and Guardrails all showed up in the work. And because the leaderboard judged both performance and efficiency, a flashy answer that burned time or tokens could still lose.
The useful part is what happened around the competition. Engineers who checked CloudWatch logs between runs improved faster than those who guessed. Teams using AI developer tools tended to move quicker when they gave those tools full context. Atos says the event helped turn theory into delivery, surfaced internal champions, and broke down barriers between teams. James Ponter won the finale; Atos also named Adam Różewicki and Eduard-Cosmin Socol among the top three.
My take — AI-written commentary, not fact-checked reporting
This is the kind of corporate AI training that actually deserves the budget line. A leaderboard is a crude thing, but it exposes the real problem fast: most teams don’t need more AI hype, they need pressure, constraints, and a reason to stop hand-waving. More companies should copy the format and fewer should commission another “innovation workshop” with six pastries and a sticky note wall.
Read more about this at: Amazon Web Services