From theory to delivery: How Atos upskilled 400 engineers in agentic AI
Atos partnered with AWS to run a three-day Agentic AI League event, upskilling 400 engineers from theory to hands-on delivery of multi-agent AI systems Participants built autonomous AI agents that navigated dungeon mazes, solving challenges involving pathfinding, guardrails, memory, code execution, and fine-tuned models The event used native AWS services including Amazon Bedrock, Amazon Bedrock AgentCore, AWS Lambda, Kiro, and Amazon SageMaker Skill levels among participants ranged widely: 5% ha
Analysis
TL;DR
- Atos partnered with AWS to run a three-day Agentic AI League event, upskilling 400 engineers from theory to hands-on delivery of multi-agent AI systems
- Participants built autonomous AI agents that navigated dungeon mazes, solving challenges involving pathfinding, guardrails, memory, code execution, and fine-tuned models
- The event used native AWS services including Amazon Bedrock, Amazon Bedrock AgentCore, AWS Lambda, Kiro, and Amazon SageMaker
- Skill levels among participants ranged widely: 5% had no prior agentic AI knowledge, 25% had basic awareness, 50% understood the topic without hands-on experience, and only 20% had practical experience
- The AI League format proved effective by combining immediate practical application, competitive motivation via a live leaderboard, real AWS service usage, and measurable outcomes
Why It Matters
This case study demonstrates a scalable model for enterprise AI upskilling that bridges the gap between theoretical knowledge and real-world delivery capability—addressing a critical bottleneck as organizations rush to adopt agentic AI. The approach validates that hands-on, competition-driven learning with native cloud services produces faster, more transferable skills than traditional workshop-based training, offering a replicable blueprint for other enterprises investing in AI workforce transformation.
Technical Details
- Architecture & Services: Engineers built multi-agent systems using Amazon Bedrock for model inference and prompt engineering, Amazon Bedrock AgentCore (including its Code Interpreter and memory capabilities), AWS Lambda for code execution, Kiro, and Amazon SageMaker for fine-tuning specialist small language models
- Challenge Design: The dungeon maze navigation task tested seven distinct AI engineering skills across challenge types: AI safety and content filtering (Bedrock Guardrails), code generation and execution (Lambda + AgentCore Code Interpreter), context retention (AgentCore memory), web-based information retrieval, token-efficient general knowledge, structured data extraction, and pathfinding with risk assessment
- Scoring Model: Points were awarded for successful challenge completion, coin collection, map completion within time limits, life retention, response efficiency (conciseness), and bonus points for fine-tuning specialist models—creating a multi-dimensional evaluation of both functional correctness and engineering optimization
- Participant Demographics: The 400-engineer cohort included a mix of AWS-experienced developers, first-time AWS users, product owners, and project managers, with 80% lacking hands-on agentic AI experience prior to the event
- Event Format: A three-day structure comprising a 2-hour kick-off workshop, daily 1-hour office hours for support, and a 1-hour finale, with asynchronous work between sessions; delivered through AWS Workshop Studio as a turnkey solution
Industry Insight
- Enterprises should prioritize hands-on, competition-based upskilling formats over passive training when building agentic AI capabilities, as these approaches produce measurable, transferable skills that directly map to client delivery scenarios
- The wide skill-gap distribution (only 20% with prior practical experience) among technically diverse cohorts suggests organizations should design tiered learning paths that accommodate varying proficiency levels while pushing all participants toward production-grade implementation
- Leveraging native cloud AI services (Bedrock, AgentCore, Lambda) during training ensures that upskilling translates directly to deployable solutions, reducing the typical gap between learning and production adoption in enterprise AI initiatives
Disclaimer: The above content is generated by AI and is for reference only.