翻訳待ち：From theory to delivery: How Atos upskilled 400 engineers in agentic AI

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

When Atos set out to upskill 400 engineers from theory to delivery in agentic AI, the team faced a familiar challenge: how to build real-world capability, not only theoretical knowledge. Online courses and classroom-based instruction build foundations, but they do not always give teams the confidence or practical experience needed to apply AI effectively to business problems. Through the Atos partnership with AWS, we had already seen that hands-on learning was the missing ingredient in effective AI enablement. We had previously delivered practical upskilling in reinforcement learning through AWS DeepRacer and in model fine-tuning through AI League in 2025. In 2026, Atos partnered with AWS to run an agentic AI League event for 400 engineers. Over three days, engineers moved from limited hands-on experience to building multi-agent systems with pathfinding, guardrails, memory, and fine-tuned models. They competed on a live leaderboard that scored both performance and efficiency. Participant skill levels varied widely. Some were developers with existing AWS experience. Others were using AWS for the first time or held less technical roles such as product owners and project managers. 5% had no prior knowledge of agentic AI. 25% had basic awareness of the topic. 50% understood the topic but had no hands-on experience. 20% had practical experience with agentic AI services. This post explains why we chose the AI League format, what engineers built and learned, which AWS services were involved, and what other enterprises should consider when running a similar event. Why the AWS AI League? Atos has a strategic commitment to agentic AI, including the development of Sovereign Agentic AI Studios in multiple locations worldwide. We needed a way to upskill our engineering teams rapidly in agentic AI patterns, not through passive training, but through hands-on delivery. Traditional workshops teach concepts. The AI League built capability through practical work under conditions that more closely reflected real delivery. The AI League format offered several advantages over conventional training: Immediate practical application – engineers built working agentic systems, rather than simply reading about them. Competitive motivation – the leaderboard created urgency and engagement that passive learning rarely achieves. Real AWS services – everything built during the event used native AWS services that engineers could apply in client delivery, including Amazon Bedrock, Amazon Bedrock AgentCore, AWS Lambda, Kiro, and Amazon SageMaker. Measurable outcomes – AI League leaderboard scores reflected functional completeness and solution efficiency. Time commitment and format AWS AI League is a turnkey solution, so setup was straightforward. Setup required only a few calls with AWS to agree on logistics and event details, plus a mechanism to advertise the event and collect participant information. Because AWS AI League was delivered through the AWS Workshop Studio, the event could be run over one to three days. To give engineers maximum flexibility, we chose a three-day format. Participant time commitments were as follows: Initial kick-off workshop (2 hours) – introduced the AI League, the AWS services used, and the format of the challenges. Daily office hours call (1 hour) – provided support for engineers who needed help or wanted to share ideas. Top three finale (1 hour) – where we crowned our 2026 champion. Outside these scheduled sessions, engineers were free to iterate on their agentic AI solutions around their existing commitments. The AI League challenge Engineers built an autonomous AI agent that navigated a dungeon maze. The agent had to find a path through the map, solve challenges on various tiles, avoid traps, and reach the treasure. All of this had to be completed within a time limit and with limited lives. The following figure shows an overview of an AWS AI League map. Figure 1: Overview of an AWS AI League map The scoring model rewarded: Successful challenge completion – correct answers earned points, while incorrect answers cost lives. Coin collection – unlike challenges, coins carried no risk and required no additional time to collect. Map completion – engineers received a treasure bonus for reaching the treasure within the allotted time. Life retention – each remaining life at the end of the challenge earned additional points. Efficiency – concise agent responses outscored more verbose ones. Fine-tuning – developing specialist small language models earned bonus points. The challenge types tested different AI engineering skills: Challenge Skill tested AWS Service(s) Violent Violet AI safety and content filtering Amazon Bedrock Guardrails Blue Brain Code generation and execution AWS Lambda, AgentCore Code Interpreter, a capability of Amazon Bedrock AgentCore Memento Context retention across interactions AgentCore memory, a capability of Amazon Bedrock AgentCore Dark Prophet Information retrieval from web sources AWS Lambda, AgentCore Code Interpreter Bonehead General knowledge with token efficiency Amazon Bedrock (prompt engineering) Healthcare API Structured data extraction Amazon Bedrock (prompt engineering) Keys & Doors Context retention across interactions AgentCore memory Spikes & Coins Pathfinding and risk assessment AWS Lambda The following diagram shows the solution architecture. Figure 2: Overview of the AWS AI League architecture Amazon Bedrock With Amazon Bedrock, engineers accessed the models that powered the agent’s reasoning. For model availability by Region, refer to Supported models by AWS Region in Amazon Bedrock: Select appropriate models for different tasks, beginning with a well-known model and then performing inference against their own fine-tuned model. Engineer effective system prompts to answer questions efficiently or delegate to sub-agents and tools. Manage token usage and cost. Amazon Bedrock AgentCore With Amazon Bedrock AgentCore, a platform to build, connect, and optimize agents at scale with any framework or model, engineers orchestrated multi-agent systems. Engineers used the following AgentCore capabilities during the AI League: AgentCore runtime, a capability of Amazon Bedrock AgentCore – engineers hosted their agent containers, which processed challenge tiles and returned scored responses within the time limit. AgentCore Gateway, a capability of Amazon Bedrock AgentCore – engineers routed tool calls from agents to Lambda functions through the Model Context Protocol (MCP) for pathfinding, web scraping, and code execution. AgentCore memory, a capability of Amazon Bedrock AgentCore – engineers persisted context across interactions so agents could recall previous events like collected keys and solved challenges. AgentCore Code Interpreter, a capability of Amazon Bedrock AgentCore – engineers executed code securely in an isolated sandbox for computational challenges. Amazon Bedrock Guardrails With Amazon Bedrock Guardrails, you can filter content to protect against harmful inputs and outputs. Engineers configured: Denied topics – specific content that had to be blocked. Content filters – thresholds for hate, violence, and misconduct. Input and output blocking with custom messages. AWS Lambda With AWS Lambda, you can build custom tool functions for tasks that models cannot reliably handle on their own: Pathfinding – navigating the map using algorithms such as Breadth-First Search (BFS). Code Interpreter – executing code for computational challenges. Web scraping – fetching and parsing web pages for information retrieval. Amazon SageMaker With Amazon SageMaker, you can build your development environment and fine-tune models. Engineers used Reinforcement Learning from Verifiable Rewards (RLVR): Amazon SageMaker Studio provided the development environment, including an integrated development environment (IDE) with built-in AI development tools. Serverless fine-tuning trained custom models on participant-created datasets. Fine-tuned models were then deployed to inference endpoints to serve traffic. The following figure shows the fine-tuning model workflow. Figure 3: Fine-tuning model overview What our engineers learned The AWS AI League surfaced several practical engineering lessons that are directly applicable to production agentic AI work. Prompt engineering under constraints One of the clearest lessons was that success did not come from writing prompts alone, but from writing prompts that worked under real constraints. Every extra token cost points. Every unnecessary tool call consumed time and cost points. Engineers quickly discovered that getting a solution to work was only the first step. Making it efficient was where the competition became most relevant to real customer scenarios. Multi-agent architecture decisions Engineers had to decide how many agents to include in their architecture, from single-purpose agents with specialist tools to more multifunctional agents. Each approach involved trade-offs in token usage, latency, and reliability. This closely reflects real production decisions about agent design. The following figure shows how different agentic systems answered challenges during the finale. Figure 4: Overview of different agentic systems answering challenges Guardrail configuration The Violent Violet challenge highlighted an important lesson: guardrails had to be precise enough to block undesirable content, without over-blocking legitimate queries. If they were too aggressive, engineers failed other challenges. If they were too permissive, they failed the guardrail challenge. This translates directly into a practical lesson in production AI safety. Pathfinding algorithm design Building the pathfinding tool required engineers to think about: Multiple strategies for different scenarios (speed versus score maximization). Risk assessment (spikes cost lives, while walls end the game). Time budgeting (should we visit every challenge, or go straight for the treasure?) Dependency ordering (for example, collecting the key before attempting the door). The following figure shows examples of multiple pathfinding strategies on the same map. Figure 5: Example of multiple different pathfinding strategies The value of observability Engineers who checked their Amazon CloudWatch Logs between runs appeared to improve more quickly. Those who guessed what had gone wrong often made slower progress. This reinforced a core engineering principle: instrument the solution properly and observe before acting. The following figure shows how to troubleshoot a Lambda function using Amazon CloudWatch Logs. Figure 6: Troubleshooting a Lambda function with Amazon CloudWatch Logs Using AI to build agentic AI solutions Engineers who used AI developer tools such as Kiro often made rapid progress. Those who shared the full context of the challenge with their AI tools tended to achieve better results more quickly. This reinforces a broader lesson: AI tools deliver more value when grounded in the specific problem being solved. Results and outcomes The event delivered: Based on the event registration data, 400 engineers gained hands-on experience in agentic AI. Engineers developed practical skills in Amazon Bedrock, Amazon Bedrock AgentCore, AWS Lambda, Amazon Bedrock Guardrails, Kiro, and Amazon SageMaker. Internal champions emerged with greater confidence in applying agentic AI in client engagements. The competitive format encouraged knowledge sharing and helped break down barriers between teams. The top-performing solutions demonstrated thoughtful engineering. They included custom pathfinding strategies, carefully tuned guardrails, memory-aware agents, and fine-tuned models designed to reduce token usage. Congratulations to our top three performers: James Ponter, Adam Różewicki, and Eduard-Cosmin Socol. Our winner, James Ponter, Head of Hype [truncated for AI cost control]