Superpower Daily
Macquarie Shifts Two Psychology Subjects to AI-Guided Learning,...Meta Introduces Muse Voice Transcribe for Live Speaker Labels i...Anthropic’s EFS Puts Misuse-Monitoring Data in Customer Clouds,...Grok 4.6 Clears 50% on Biosecurity Refusal and Routine Biology...HUMAIN Picks Compal for Cloud-to-Edge AI Build, With Nvidia and...t54 Puts a Hard Trust Gate on Bedrock Agent Payments, Citing 20...Macquarie Shifts Two Psychology Subjects to AI-Guided Learning,...Meta Introduces Muse Voice Transcribe for Live Speaker Labels i...Anthropic’s EFS Puts Misuse-Monitoring Data in Customer Clouds,...Grok 4.6 Clears 50% on Biosecurity Refusal and Routine Biology...HUMAIN Picks Compal for Cloud-to-Edge AI Build, With Nvidia and...t54 Puts a Hard Trust Gate on Bedrock Agent Payments, Citing 20...
Businesspublished

Atos and AWS Put 400 Engineers Into a Scored Multi-Agent AI Contest

The dungeon challenge made engineers balance route planning, memory, tool use and safety controls against time, lives and response efficiency—not simply produce a working answer.

By 2 min read
Atos and AWS Put 400 Engineers Into a Scored Multi-Agent AI Contest
Atos and AWS Put 400 Engineers Into a Scored Multi-Agent AI Contest
Sign in to save and react

Listen to this story

The audio brief

About 1:43
0:001:43
Read transcript
Atos and AWS turned AI training into a three-day engineering contest, with 400 people building agents that had to survive a dungeon maze—not just produce a correct answer. Each team’s system had to choose routes, solve challenges, avoid traps, and reach treasure before time or lives ran out. The leaderboard also rewarded accuracy, concise responses, low token use, low latency, and reliability, making efficiency part of the engineering problem. The starting line was uneven. Atos reported that 5 percent of participants had no prior knowledge of agentic AI, while half understood the ideas but had no hands-on experience. Another 20 percent had already used agentic AI services. The maze gave all of them a concrete way to confront the tradeoffs enterprise deployments face: whether to use specialized agents or more multifunctional ones, how much context to retain, and when guardrails become so strict that they block legitimate requests. Under the hood, AgentCore Memory preserved keys and solved challenges across interactions. AgentCore Gateway used MCP to connect pathfinding, web-scraping, and code-execution tools running on Lambda. Engineers worked across Bedrock, Guardrails, Kiro, and SageMaker, with bonus points available for fine-tuning small language models. Atos says the event created internal champions for client work and supports its AgentX framework for incident triage, investigation, resolution, and IT service-management updates. AWS says enterprise AI League events are available through 2026. The open question is whether contest-built skills translate into measurable production outcomes—the story provides no deployment metrics yet.

Story brief

3 key points

Atos used a three-day AWS-backed AI League to move 400 engineers from classroom concepts into a scored multi-agent systems exercise. Teams were judged not just on reaching treasure but on accuracy, lives, response concision, token use, latency, and reliability, exposing the engineering tradeoffs enterprise deployments face. The contest also served as a pipeline for AWS tooling—Bedrock, AgentCore, Lambda, Guardrails,...

  1. 01

    Participant readiness varied: 5% had no agentic-AI knowledge, 50% understood it without hands-on experience, and 20% had practical service experience.

  2. 02

    AgentCore Memory preserved keys and solved challenges; Gateway used MCP to connect pathfinding, scraping and code-execution Lambda tools.

  3. 03

    Teams balanced specialized versus multifunctional agents while managing token consumption, latency, reliability, guardrails and legitimate-query blocking.

Atos and AWS ran a three-day AI League for 400 Atos engineers, turning agentic AI training into a live contest for building multi-agent systems. Participants made autonomous agents navigate a dungeon maze, solve tasks, avoid traps and reach treasure within time and life limits; the leaderboard measured performance and efficiency.

The event drew people with sharply different starting points. Atos reported that 5% had no prior knowledge of agentic AI, while 50% understood the topic but lacked hands-on experience. Another 20% entered with practical experience using agentic AI services.

The maze turned agent design into a set of choices

Participants built agents to choose routes through tiles containing coins, traps and challenges. Correct answers earned points, while wrong ones cost lives; reaching the treasure and keeping lives raised a score. The work covered multi-agent coordination, pathfinding, web retrieval, code execution, memory, prompt engineering, guardrails and model fine-tuning.

AWS AI League maze interface showing a grid with walls, coins, traps, challenge tiles and a treasure chest.
The maze combined navigation, task completion and risk decisions in one simulated workload. Source: aws.amazon.com.
Participant starting points
5%No prior knowledge

Atos reported that 5% of registrants had no prior agentic AI knowledge.

50%Understanding, no hands-on work

Atos reported that half of participants understood agentic AI but had no hands-on experience.

20%Practical experience

Atos reported that 20% had practical experience with agentic AI services.

Efficiency was a first-class requirement

A working answer was not enough. Concise responses outscored verbose ones, and participants could earn bonus points by fine-tuning specialist small language models. Engineers also had to choose between more specialized or more multifunctional agents, weighing token use, latency and reliability.

Three technical pressure points

  • Guardrails: engineers configured denied topics and content filters, while trying not to block legitimate queries.
  • Memory: AgentCore memory retained context across interactions, including collected keys and solved challenges.
  • Tool use: AgentCore Gateway routed calls through MCP to Lambda functions for pathfinding, web scraping and code execution.

Training for a broader AWS agent push

Engineers used Amazon Bedrock, Bedrock AgentCore, Lambda, Bedrock Guardrails, Kiro and SageMaker. Atos says the event created internal champions and increased confidence in applying agentic AI to client engagements.

The exercise sits alongside Atos’s AgentX, which the company describes as a multi-agent framework powered by AWS agents and purpose-built agents for cloud operations. Atos says AgentX is designed to handle incident triage, investigation, resolution and IT service-management updates.

AWS says the AI League format is available for enterprise events through 2026, as well as at select AWS Summits and virtual events.

Sources

  1. aws.amazon.comFrom theory to delivery: How Atos upskilled 400 engineers in agentic AI | Amazon Web Services
  2. atos.netAtos AWS AgentX platform

Related coverage

Continue reading