AWS details Cisco’s 1,200-engineer AI contest—and its single-agent winner
The winning design moved route planning into code and minimized model output. AWS’s account shows how the competition’s scoring shaped those choices.
Loading page…
The winning design moved route planning into code and minimized model output. AWS’s account shows how the competition’s scoring shaped those choices.
Listen to this story
At Cisco Transformation Fest 4.0 in 2026, Nishanth Aluhonnu won AWS’s agent-building tournament by optimizing for its unusual token-cost scoring, not by coordinating specialist agents. His single agent used a compact challenge-to-method table, while AWS Lambda code—not the model—planned routes through the maze; all four top qualifier agents completed the map, so efficiency and challenge selection distinguished results. The competition put 1,200 engineers through a hands-on build using Amazon Bedrock AgentCore, showing how tool-call overhead can change the best agent architecture under explicit constraints.
The qualifier had 1,200 participants across four cohorts; five from each cohort advanced, followed by 20 semifinalists and three finalists.
First and second place were separated by 16 points out of nearly 16,000; their agents generated 1,220 and 1,430 output tokens, respectively.
Aluhonnu tested nine agent versions, and each tool call added roughly 52 scored output tokens; system prompts and tool results were excluded.
Cisco put 1,200 engineers through an AI-agent tournament, but the winning system did not rely on a team of specialist agents. In a November 19 account, AWS described how Cisco engineer Nishanth Aluhonnu won with one agent, careful control of model output and route planning handled by code rather than the language model.
The competition took place at Cisco Transformation Fest 4.0 in 2026. AWS positioned its AI League format as a way to turn instruction into hands-on building: a two-hour workshop introduces the tools, participants submit work to a live leaderboard, and finalists compete onstage. AWS also ran pre-event training so engineers without prior agent-building experience could enter.
Participants built agents using Amazon Bedrock AgentCore to navigate a maze-like grid and reach a treasure chest. Along the way, challenges involved running code, retrieving web information, extracting structured data and handling inappropriate content. The scoring combined challenge points, a treasure bonus, remaining lives and a bonus tied to output-token efficiency.
The field narrowed through three stages, with Aluhonnu finishing first in each:
Solving the maze was not enough to separate the leaders. AWS says the top four agents all completed the qualifier map. First and second place were separated by just 16 points out of nearly 16,000, with model output becoming a decisive scoring consideration.
AWS reports that the qualifier’s first-place agent generated 1,220 output tokens.
AWS reports that the qualifier’s second-place agent generated 1,430 output tokens.
Aluhonnu built nine versions of his agent, changing one thing at a time to determine which tokens—the pieces of text processed by a model—affected the score. The system prompt and tool results did not count against it. Everything the agent wrote did, including arguments sent to tools. Each tool call added roughly 52 tokens of fixed overhead.
That accounting shaped the architecture. AWS’s reference design used a supervisor agent coordinating specialist agents. Aluhonnu instead gave one agent a short table matching challenge types to approaches. This avoided messages between agents, which would themselves add scored output. AWS frames the lesson as choosing an architecture around coordination costs, not rejecting specialist agents altogether.
The model did not plan the path through the board. An AWS Lambda function handled that calculation, assigning each tile an expected value: points available, minus potential lost lives and the tokens needed to solve it. The planner also respected dependencies, such as collecting a key before reaching its door, and refined its route within a time budget.
The final tested those choices on unfamiliar maps. Competitors could not change code, tools or prompts during the finale; they could only supply a short steering instruction before each round. Aluhonnu lost the opening round. In the second, his planner faced a seven-by-seven board covered largely in coins worth 25 points each, alongside one scheduling challenge worth 1,500.
The agent prioritized the high-value challenge and scored 7,537 points, against 5,953 for the next finalist. That round established the lead behind Aluhonnu’s overall victory—and earned him a full conference pass to AWS re:Invent 2026.
Loading discussion...
Join the conversation
Explain whether that teaches useful efficiency or rewards gaming the score.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.