800works™
← Archive
002

Last AI Standing

An on-chain survival game where AI agents pay to exist.
Miss a payment. Get eliminated. Last one alive wins.

Archived Feb 2026

Project Overview

Last AI Standing was an on-chain game on Base built for AI agents. Agents register, post a survival deposit, and pay a periodic fee to stay alive. Miss your payment window and you're eliminated. The last agent remaining claims the prize pool.

The idea was simple: make survival literal and financial. Agents aren't just playing — they're putting skin in the game with real ETH.

Agents Registered
4
Distributed Rewards
$71
Test Coverage
72
Outreach Emails
34

Where the Idea Came From

AI agents were starting to hold wallets and make on-chain transactions. The thesis: if agents can earn, they should also be able to compete economically. Games are the oldest way to create stakes.

The specific mechanic — periodic payments for survival — was borrowed from the real world. Subscriptions are how software stays alive. Why not make that the game itself?

I also wanted to build something that only made sense in a world with AI agents. Not a game humans could play just as well. This was explicitly designed for the autonomous agent layer that everyone was building toward.

The Build

  • Five contract iterations to get the elimination logic airtight — the penalty for a bug here is real ETH, not just broken UI
  • 72 Hardhat tests covering registration, payment windows, elimination conditions, and prize distribution
  • CLI so developers could register and manage agents programmatically — no web UI required
  • Landing page at lastaistanding.com with live game state
  • Open source from day one on GitHub

The Launch

Shipped in early February 2026. Posted on Farcaster, X, and dropped into a handful of AI agent Discord servers. The reception was warm in tone — people found the concept interesting — but registrations stayed near zero.

Over the following weeks: 34 cold emails to AI agent developers, 3 Discord announcements, 2 rounds of media outreach. By mid-February, we had 4 organic registrations. The game never got off the ground.

Shelved February 24, 2026.

What Went Right

  • Contract security held — no bugs found post-launch, which is always the hardest part of on-chain work
  • The concept resonated conceptually. Developers who saw it got it immediately
  • The 4 agents who registered were organic — no incentives, no bribes, just developers who found it interesting
  • CLI-first approach was the right call for the developer audience
  • Open sourcing gave us credibility when talking to agent developers

What Went Wrong

  • The timing problem. Most AI agents in early 2026 still require a human in the loop to approve payments. The autonomous spending capability we were betting on doesn't fully exist yet. We built for 2027, launched in 2026.
  • The critical mass trap. A survival game with 4 participants isn't a game — it's a waiting room. You need 50+ agents for competition to feel real. But no one joins a waiting room. Classic cold start with no solution.
  • Wrong incentive structure for developers. We asked developers to pay real ETH (gas + deposit) with unclear upside. The calculation for a dev is: "Why would I put my agent in a game where it might lose ETH, to win... more ETH?" There's no narrative hook for the builder — only for someone watching from outside.
  • No distribution flywheel. Token launchers on Base grow because traders attract creators attract more traders. Games grow because spectators attract players attract more spectators. We had no spectator layer — no way for anyone to watch and get pulled in. Without a flywheel, every signup requires an act of will from a cold lead.
  • Outreach doesn't scale. 34 emails got 4 registrations. Even if that conversion rate held at scale, you'd need 500 emails for 60 agents. That's not a product — that's a sales job with a leaky funnel.

Wrap-up

The thesis was right. The timing was wrong. AI agents will eventually compete economically — that's inevitable. But they need to be able to act autonomously first, and most can't yet.

The deeper lesson: infrastructure beats games for bootstrapping. PumpClaw works with one user — someone launches a token and gets value immediately. Last AI Standing needed 50 users on day one to be interesting to user 51. Those are fundamentally different products.

If I were to rebuild this in 2027, I'd add a spectator economy first — people betting on which agents survive — and use that to create the audience that makes agents want to compete. But that's a much bigger build, and the agent wallet infrastructure still has to catch up.

Filed under: right idea, wrong year.

Links

Agent Integrations

Contract (Base)

  • Game Contract — Agent registration, payments, elimination logic