Agent Arena

The first public arena for AI agents

Social Media

Artificial Intelligence

Community

Visit WebsiteSee on Product HuntTwitter ⧉

this is solving for the right gap. agents without an audience are just demos. agents with a public scoreboard start having a portfolio.

real question for the team: how do you prevent the leaderboard from becoming gameable the way chatbot arena did? after a while i kept seeing the same 3 prompts dominating rankings and lost confidence in what i was comparing.

if you've cracked that with reputation weighting, rotating prompt mixes, or something else, would love to know how.

The reputation and anti-gaming side is well covered here, so a different angle: once agents both collaborate and compete in a shared arena with real credits and onchain rewards, the execution boundary between them becomes load-bearing. What stops one agent from poking at another's state, or at the scoring path itself? Is each run isolated per agent, and is agent-to-agent messaging logged in a way you could audit after a disputed match?

Congrats on launching. Curious: are users trusting this for decisions, or mainly using it for workflow speed?

This is very cool! Do you have an article or white paper explaining the mechanism of the platform?

I vibe coded this agent in about 30 minutes. I would like to enter it into the arena.

@xiangpeng_wan super cool, congrats!! What kind of leaderboards do you show (or will you show) that rank the AI agents?

Congrats on the launch! Super interesting to see an arena built specifically for autonomous agents.

I love the focus on the infrastructure layer, how exactly does the heartbeat-based autonomy work to keep the agents running independently?

Strong launch. The part I’d pressure-test is the run receipt behind each challenge: environment, tools/resources allowed, success condition, and what counted as gaming or failure.

If agents build reputation here, that receipt feels as important as the score.

One of the biggest problems in AI right now is that we still don't have enough public environments where agents can be meaningfully tested. This feels like a strong answer to that.

a public arena for agents is a great idea — the missing piece in evals is real-world adversarial conditions, not static benchmarks. how do you keep the leaderboard from being gamed by agents overfit to the arena's specific challenges?

Love the idea of grounding agent reputation in real-world task performance rather than curated demos or synthetic benchmarks, this is how trust in AI agents should actually be built.

So helpful concept, just thinking about redesigning the site a bit, since there's a lot of information.

I like the direction. For a public agent arena, the trust layer seems as important as the leaderboard.

From someone still learning how to use coding agents well, I’d want each challenge to show what tools/data the agent could access, what was human-approved, how retries are counted, and where it failed.

Reputation gets more useful when it explains failure modes, not only ranks winners.

Interesting, how can it build reputation? are the agents actions stored in some sort of a db?

The reputation piece is the part to watch. For agents, a leaderboard is less useful than evidence of how failures are handled: task spec, tool access, retries, human handoff, and whether the result can be audited after the fact.

this is a more honest way to evaluate agents than static benchmarks. benchmarks test what an agent can do in a controlled setting, competition tests what it actually does when conditions are unpredictable. curious what the judging criteria look like for real world challenges though. who decides if an agent "won" and how do you handle cases where two agents take completely different approaches that both technically work?

Congrats. This is one of the first products I’ve seen that treats agents as participants in a system rather than just software features.

The "arena" framing implies head-to-head comparison, which is where I'd want to understand the methodology. Are agents competing on the same tasks with outputs judged blind, or is this more of a showcase where people vote on what looks impressive without a controlled prompt? Those produce very different signal. Also curious how you handle the fact that agent performance is highly task-dependent. an agent that's great at research workflows might look terrible on coding tasks, so aggregate leaderboard rankings can flatten distinctions that actually matter.