A November 19 account by AWS and Cisco participants says 1,200 Cisco engineers competed in an agent-building tournament at the fourth Cisco Transformation Fest. The AWS account team provided training ahead of the event. Four qualifying cohorts sent five people each to a semifinal, and three reached a live final. Cisco engineer Nishanth Aluhonnu won and received a pass to AWS re:Invent 2026. The numbers show an organisation getting a large group to try a defined technical task. They do not say that all 1,200 can put an agent into production. The decision worth examining is how a company pays for enough practice to give staff an informed starting point. The Cisco and AWS case
My view is that Cisco’s format is worth copying when an employer can give people time to build and a useful problem to return to after the game. Participants get a reference point for the performance of their work, see alternatives and hear other people explain their decisions. The value can continue if enough of them can use that knowledge without relying on the event’s prepared environment. A manager should budget for a second phase and name the kinds of work the training is supposed to improve. The first sign of success is that engineers can make and explain a decision that they previously needed someone else to make for them.
Put practice and feedback in the same room#
AWS describes the AI League format in three stages: a workshop, a period to build and submit, and a live final. The general programme describes a two-hour introductory workshop. In Cisco’s contest, participants faced a simulated problem with a scoring system and had opportunities to submit work and try again. The benefit is the short distance between making a choice and seeing its consequence. A person can change one aspect of a solution, get measurable feedback and discuss why a different approach behaved better. That cycle is difficult to run in a passive session even when the speaker has good examples. It helps explain why a company would allocate working time to the format. AWS AI League programme
There is a second organisational benefit to a common task. Two independent engineering groups may use the same word for different problems. One team may think of a model as a way to understand a text. Another may expect the model to make a decision about which step comes next. A shared task exposes those assumptions in a way a slide cannot. If a group compares where each approach ran out of time or produced the wrong result, it has something concrete to discuss. A manager can choose a challenge with enough room for different approaches and use the shared feedback to make participants’ choices understandable across teams.
That structure can help beginners even if they never see the final. The case says Cisco’s advance training was intended to let engineers without previous hands-on agent experience compete. A beginner might finish a task by following an example, then find out what happens when they alter a constraint. That is a modest and meaningful result. A programme should be able to show how many people could complete and explain the basic work on the way out. A winner’s score tells an interesting story about what is possible under the rules. It gives a manager much less evidence about whether the rest of the workshop left people better prepared to do useful work.
A leaderboard changes which decisions people notice#
The tournament evaluated aspects including accuracy, time, planning and output token use. In the case report, the winner tried nine variants and examined how the scoring rules counted model text and tool arguments. Other possible approaches used several agents to divide the task. The winner used one and assigned some deterministic work to a tool. This is interesting because it makes an engineering decision visible. Coordination can consume time and output, and a simpler structure can work when the problem and measurement reward it. The report does not establish a universal rule about which design is best. It establishes why design and measurement belong in the same conversation.
The final also restricted what the participants could change during the contest. The report says they could give a short steering instruction before each round, but could not change the code, tools or prompts. The agents had to perform on maps they had not seen. That adds a useful kind of variation to training. Participants had an incentive to make a solution cope with new conditions instead of making it work for just one known example. A company taking the exercise into a real department can do something similar by retaining an example for evaluation and deciding beforehand what an acceptable response looks like. The point is to see whether a decision transfers to a case that gave the author no special advantage.
Choice of measures also changes what people build. In the case, output text affected the competition score in a particular way. A production system gets bills for a different set of resources and needs acceptable results when a user cannot tell whether the agent was making things up. A group bringing work from the game into an application should restate the whole set of conditions. It might need to measure total cost, user time, quality on representative inputs and the work required when something goes wrong. A deliberately simple competition rule can teach people to look for trade-offs. Production measures should then use the trade-offs the actual company has to make.
Participation is a useful measure with a limited job#
Cisco has a useful case about scale. A group of 1,200 showed up to a contest in four cohorts, while the account describes the broader annual Transformation Fest as an event reaching around 2,000 engineers. Those figures have different meanings. They do not add up to a new participant total, and they do not indicate how many people were able to repeat the work after the event. A report can show how many registered, tried the task, submitted something that ran, later tried another case, and completed supervised work. Keeping the stages visible makes participation more valuable. It gives managers a starting denominator and a way to see where the next investment should go.
There is a historical comparison inside the programme. In 2025, Cisco engineer Hemanth V first competed in an internal model customisation challenge and later won the agent challenge at AWS re:Invent. In 2026, Nishanth won the Cisco event and earned a conference pass. Those are different outcomes at different points in the programme. They show what can happen when staff get more chances to practice and compare with peers. They say much less about average competence. A credible report would include direction for people who were never ranked near the top, since they may be the ones who move a training idea into ordinary team practice.
The earlier AWS and Atos account gives another version of the competition format. It describes 400 participants in a 2026 event, with a range of starting experience and a mix of scheduled support and time to iterate independently. The case comes from another participating organisation and the same training provider. It gives context for how the contract with participants can vary. It does not supply a control group for Cisco’s outcome. An organisation comparing programmes can ask whether its students need access to a facilitator, how much practice time they can take and what the next test asks them to do. The comparison is more useful when it tracks those conditions. AWS and Atos’s earlier account
The budget pays for practice time as well as tools#
An AI League event benefits from an existing partnership and prepared exercises. AWS says enterprise hosts can apply for credit for an internal competition. The economics extend beyond AWS credits. An engineer needs time away from ordinary delivery, and a mentor needs time to help people who get stuck. Managers may also have to prepare approved test material, ensure accounts work and provide equipment or access across time zones. That is part of what the company buys when it supports practical learning. A credible budget tells supervisors how much time they can assign and protects some of it. A training programme for a large group will struggle if the work is only possible in a few unclaimed hours.
Different groups will incur different costs. An engineer who uses cloud tools daily can start experimenting fairly quickly. Someone who manages deployment in another environment may need orientation before they can interpret an error or know which settings they should touch. The programme can offer work of several levels and count help requests as useful information. If the only public measure is rank, people who have the most to learn may drop out early. A group that offers office hours and space to discuss unsuccessful attempts can discover what was difficult to learn. It is still paying for technical capability, even if part of the intervention looks like ordinary employee support.
The choice of cloud is also an organisational choice. A vendor’s services can make the first practice easier, particularly where teams already use them. Skills become more durable if participants can explain which parts belonged to the specific service and which decisions would also apply elsewhere. A good example is the choice of what to measure, when to ask for help or when a smaller design works. Another organisation that buys or builds on a different platform should check whether a participant can transfer those decisions before assuming that a competition took care of training. A workshop can be a very useful introduction above the product-specific layers. The employer still needs to say which parts it wants to retain.
Give people a second task under ordinary team conditions#
The contest should be followed by a task that is small enough to review and useful enough to have an owner. A team might want a tool to summarise approved internal documents, with a way to compare the output with the source. It might want help preparing a draft that a person always inspects before sharing. Either case should have a written boundary for data and intended audience, an ordinary path to run the test and someone responsible for deciding whether it is good enough. The second exercise is where the company can learn whether people recognise the assessment questions that were implicit in the game. It also gives colleagues a chance to reuse work even if they did not attend.
The labour division from a competitive environment can change on a real project. A developer may have found that a single agent beat several in the tournament. A team deploying to a product might still need a division of responsibility for data access, review or changes that affect other staff. It can also discover that a regular program does one part of the work more reliably than any model. Those are respectable outcomes for technical training. The test is whether someone can explain the role assigned to the model and what evidence supports it. Training makes a stronger case for itself when people can bring a design principle to a project and decide against using a feature that adds little value.
An industry example does not tell a manager which business outcome to aim for. It can show how to make the learning process specific. An owner might need same-day drafts during an approved weekly process. Another may need fewer errors in an existing review, without changing when it is delivered. Comparing the old and new methods against one goal prevents a small reduction in typing time from being mistaken for a faster delivery. It also makes it easier to tell what to fix after a poor result. A basic handoff note and a person who can repeat the exercise add more lasting value than a diagram of a promising prototype.
Look for sustained capability after the event#
One possible measurement is a comparison of unfamiliar but comparable tasks before and after the event. Use the same quality criteria and record any support given. A participant can demonstrate how they evaluate the result, stop when it is wrong and agree a next step with a teammate. A second check after a few weeks can show whether they still understand the choices. It is important that people have enough time to take part and that the organisers do not confuse familiarity with the event environment and a transferable skill. A useful result might be that more participants can repeat the work without assistance while explaining how to detect an error.
The later review can follow a few outputs into the organisation. Which team tried a new idea? What were the intended users and first acceptable results? Which experiments were stopped because the approach failed under real conditions? Which would require a different training request to move forward? Count finished pilots against the work needed to support them, with the same distinction between training results and an outcome seen by users. An organisation that improves at stopping poor candidates may gain value without adding many applications. This should be visible to managers deciding whether to fund another event or give existing participants more review time.
There are plenty of good reasons for a broader programme to continue even if a particular pilot fails. Cross-team contacts may make it easier to find someone who has tested a service or understands a limitation. Training materials can give a group a shared vocabulary and reusable measurement. Those benefits are easier to claim with tangible records: a colleague using the old notes, a manager borrowing a comparison exercise, or a solution that a second team could explain and run. A local programme can even share evidence back with the provider about where the exercise does and does not resemble work. None of that requires a universal headline about mastery.
Make the learning available to people who never reach the final#
A final is effective at giving people a memorable example. The organisation receives more value if it extracts several examples and leaves time to discuss them. One useful account can follow the winner’s reasoning and ask which rule encouraged it. Another can follow someone who completed a simpler task and explain what they needed to learn before their attempt ran. Both help a future group decide where to start. A short record of differences in tools, assumptions and checks is more durable than a list of points. The point of preserving ordinary results is that engineers who join the next round can begin with the discoveries that are relevant to their starting level.
The people doing the follow-up should be named before the event. A lead can request two short sessions from the team that designed the exercise. Another manager can assign time to turn a promising idea into a written, reviewable trial. A support person can help when a new team uses an example on a different set of allowed documents. None of those roles depends on running the same game again. They require a clear place for materials, working time and answers to questions about access. An organisation can budget for these activities deliberately and measure participation by the people who use the resulting practice, not only those who read an invitation.
There is also a reason to preserve the cases that did not finish. Participants unable to start may have lacked time to set up the environment, a familiar way to check the result or enough background to understand the terminology. Others may have had time to begin and discovered too late which outputs mattered. Those are distinct problems that suggest different uses of the next budget. A more useful orientation may help one group, while an example with immediate feedback may help another. Taking missing completions seriously allows the programme to improve for future participants who would not otherwise be evident in the final. A large attendance figure creates enough variety to make this worth examining.
Compare the programme with other ways to use the same time#
The strongest managerial comparison uses the same quantity of staff time. A group could spend a day in a supported practical exercise, build around a real project with a mentor or take more small practice sessions spread across a month. Each has a different value for people who start with different experience and current responsibilities. A contest is good at providing a common deadline and immediate feedback. It can lead to performance under conditions unlike a normal week. Mentored project work can produce an output a team immediately uses, while reaching fewer people at once. Asking what the learners are able to do later makes those options comparable before selecting another format.
Managers can also check which benefit would justify the expenditure. If teams could already make a good initial prototype but struggled to agree on quality, another competition may add less than shared evaluation time. If the obstacle is a lack of accessible experience, a wider exercise could give more people an entry point. Record the baseline and try to find a second independent measure of improvement, such as whether a colleague can reuse a participant’s reasoning or reproduce a test. This does not require a formal cost study for every lesson. It requires enough evidence to defend the time spent and make a better decision the next time.
Put the programme’s next decision on the calendar#
An organisation running an event can set a date when leaders decide whether to repeat, narrow or change it. At that meeting, it is useful to compare what the group learned with the changes that made it into actual work under acceptable conditions. The people who own the first production trial need a known support path, and mentors should be able to say what skills still required help. If the trial creates more review overhead than usable output, the next expenditure might be in evaluation or mentoring. If a second set of engineers can repeat and improve an approved job, the programme may have earned more time and a larger audience.
Cisco’s reported participation and successive tournament results are substantial signs of engagement. They make a reasonable case for practice supported by a large employer. The company’s long-term return will depend on whether a much wider group can reliably recognise a task, choose an appropriate tool, test what it does and take responsibility for the final work. Those abilities can be assessed on ordinary projects without claiming that a game has already measured them. The people at the top of the scoreboard have done something concrete. The next month will show how much of that opportunity can be shared.