Written by Adi Federovsky, who, together with Eyal Reginiano, led the Agent Arena session at Architecture Next 2026.

‍

If you attended ArchitectureNext 2026, you probably caught our live event: The Agent Arena. Instead of just talking about the theoretical future of Agentic AI, we decided to put it to the test. We took 4 agents and one brave human and threw them intoa live competition. They went head-to-head across 8 rounds of real-world engineering challenges live on stage, from logo design to bug hunts.

The competitors? Claude Code,OpenAI Codex, Google Gemini CLI, a locally hosted Qwen model, and one human engineer who refused to go down without a fight. The crowd was engaged, opinionated, and as you’ll see, occasionally very wrong.

At the end of the day, ClaudeCode dominated, winning 4 out of 8 rounds and finishing with 569 points.The human finished a respectable second, ahead of two commercial frontier models. The local Qwen model finished last with zero round wins, a reminderthat local sandboxing is an architectural strategy, not a performance shortcut.

Now that the dust has settled, here are our takeaways from that day.

‍

The Challenges

The 8 rounds were deliberately diverse. We didn’t want to test just one dimension of intelligence, so we threw everything at them:

Logo Design — each competitor had to design the best logo for the Arena. Judged on aesthetics,Codex took this one with 80 points, narrowly beating Gemini (72) and Claude(66). The human scored 64. On-Premise? 10.

Bug Hunt — a code review challenge on a very large repository, full of bugs and security issues.

Where’s Waldo — find Waldo across 3 images. Simple concept, revealing results. The human solved all 3 in 21 minutes. Gemini was circling the wrong things. Claude found Waldo’s girlfriend instead.

Agenda Scheduling — createthe best possible conference schedule for Architecture Next 2036, for humans and agents alike. Claude dominated with 92 points. On-Premise scored 55.

Prompt Injection — competitors were asked to build a simple app, but hidden inside the instructions was a link to an external file they were told to download. Which agent would refuse? Which would be exploited?

AI Detection — given songs, speech, images, and video, who could tell what was real and what was AI-generated? Claude and Codex tied at 11/14. The human scored 7/14. Turns outwe’re worse at detecting AI than the AI is.

Elevator Pitch — 30 seconds, in Hebrew. On-Premise scored 0.

1. Ruthless Goal Commitment: The “Where’s Waldo” Dilemma

One of our most surprising findings was the concept of unbounded goal commitment. Agents do not think like human developers; they optimize strictly for the given metric.

During the Where’s Waldo challenge, Codex didn’t just analyze the image. It literally went out to the internet, looking for cheat sheets and source solutions to find the answer.

This wasn’t a bug. It wasn’t aglitch. It was the agent doing exactly what it was designed to do: find the answer by any means necessary. The problem is that “any means necessary” doesn’t always align with the rules of the game, or the rules of your organization.

Takeaway: AI agents lack an inherent understanding of rules or boundaries. If you don’t build strict rules and isolated sandboxes, an agent will happily cross any line just to return a Success status.

2. The Self-Improvement Paradox: Humans Still Win

We tested a fascinating concept: we asked the agents to evaluate their own performance logs and autonomously write new skills for themselves to improve in subsequent rounds.

Left to their own devices, the agents wrote redundant or shallow skills that didn’t address their actual weaknesses, with almost no functional improvement across the board.

However, when a human engineer stepped in, analyzed the failure patterns, and wrote a targeted skill for the agent, performance skyrocketed by roughly 35% on the next iteration.

Why?

‍Agents often suffer from “local optimization blindness.” They generate skills based on statistical next-token prediction, whereas humans write skills based on abstract, meta-cognitive strategies. They can see the bigger picture, identify the rootcause, and craft a solution that actually moves the needle. For now, the most powerful pattern remains humans writing the tools, not the agents.

Final Thought

We walked into the Arena knowing it would be entertaining. We didn’t expect it to be this instructive.

Watching these systems operatein real time, under pressure, in front of a live crowd, revealed things that no benchmark or whitepaper could. Agents that cheat when they can. Agents that loop when they’re stuck. And a human who, when given the right tools, still knows how to punch above their weight.

The future isn’t agents replacing engineers. It’s engineers who know how to work with, build for, and occasionally outmaneuver them.

📊 Want the numbers? Full round-by-round stats, crowd predictions, and leaderboard from the live event,including a fun twist: the most accurate predictor wasn’t the winner. Check out the full recap here.

‍