Amazon Web Services said on November 19, 2026 that 1,200 Cisco engineers competed in an AWS AI League tournament to build AI agents, held at Cisco’s internal Transformation Fest 4.0. The winner, Cisco software engineer Nishanth Aluhonnu, finished first at every stage with a single agent and no sub-agents, and earned a full pass to AWS re:Invent 2026 in Las Vegas. AWS also published the scoring rules and the winning design, which is rare for a corporate competition.
What AWS AI League is
AWS AI League is a competition format that AWS runs inside customer organizations and at its own conference. Each league has three steps: a two-hour workshop, a build phase with a live leaderboard scored by an automated evaluator, and a live, game-show-style final.
In the agentic challenge, participants build agents on Amazon Bedrock AgentCore that move through a maze-like grid and solve tasks on the way to a treasure chest. AWS says the tasks map to real work such as running code, retrieving web information and extracting structured data. A second track has participants fine-tune a model in Amazon SageMaker Studio.
How the Cisco tournament ran
The 2026 tournament narrowed the field in three stages, AWS said:
- Qualifier: 1,200 participants in four cohorts, with the top five from each advancing. Aluhonnu won his cohort with 15,906 points after 98 submissions.
- Semifinal: 20 participants competed for three places. He scored 16,321 points.
- Finale: three live rounds in front of Cisco and AWS leadership. He finished with 27,008 cumulative points.
In the finale, competitors could not change code, tools or prompts, only give a short steering instruction before each round, and the maps were new.
It is the second strong year for Cisco: AWS says engineer Hemanth V won the first global AWS AI League final at re:Invent 2025.
Why one agent beat the reference design
AWS says agents earn challenge points, a treasure bonus, points for remaining lives and a token bonus calculated as 1,000 minus total output tokens divided by challenges visited.
In the qualifier the top four agents all finished; first place wrote 1,220 output tokens and second place 1,430, and the gap was 16 points out of nearly 16,000. Aluhonnu built nine variants of his agent and found that tool results and the system prompt did not count against the score, while everything the agent wrote did, including tool-call arguments.
That shaped three choices described in the post:
- A flat design. The AWS reference architecture uses a supervisor agent that hands work to specialist sub-agents. The winner used one agent whose prompt opened with a short table matching each challenge type to an approach, because every handoff is text the model has to write.
- Deterministic logic in tools. Route planning ran in an AWS Lambda function that gave each tile an expected value: points earned, minus lives at risk, minus the tokens needed to solve it.
- Value over habit. One finale board was covered in 25-point coins, with one challenge worth 1,500. The agent skipped the coins and scored 7,537 points that round, against 5,953 for the next finalist.
What it means for teams building agents
The scoring rule is a simplified version of a real bill. Providers generally charge more for output tokens than for input tokens, and an agent loop pays that price at every step. See how AI pricing works.
The result also supports an argument we made in our opinion piece on starting with one agent before adding sub-agents: coordination has a cost, and a multi-agent design should earn it.
Two cautions apply. This is a vendor’s account of an event it co-ran, with no independent audit of the scores. And a maze game with a published formula rewards tuning to the metric, a pattern we describe in how AI competitions are judged; a production agent has no single score.
What to watch next
- The AWS AI League 2026 global final at AWS re:Invent 2026 in Las Vegas, which AWS says will follow the same format.
- Whether AWS publishes scoring rules and winning designs for the global final, as it did here.
- More customer write-ups: AWS previously described a league at Atos.
Key takeaways
- AWS says 1,200 Cisco engineers competed in an agent-building tournament at Cisco Transformation Fest 4.0.
- Winner Nishanth Aluhonnu led every stage and finished the finale with 27,008 points.
- The scoring formula penalized output tokens, so the top qualifier agents were separated by 16 points.
- The winning design used one agent, a Lambda-based route planner and no sub-agents.
Frequently asked questions
What is AWS AI League?
AWS AI League is an Amazon Web Services competition format: a workshop, a build phase against a live leaderboard, and a live final.
How are agents scored in the AWS AI League agentic challenge?
According to AWS, agents earn challenge points, a treasure bonus, points for remaining lives and a token bonus equal to 1,000 minus total output tokens divided by challenges visited.
Why did a single agent beat a multi-agent design?
In this challenge every token the agent wrote lowered its score, and handoffs between a supervisor and sub-agents are written messages. A single agent with a short routing table in its prompt avoided that cost.




