Josh SolisJS Book an intro callBook a call

AI agents · Marketplaces · Infra

Agent Systems: AI agent platforms: escrow, spend limits and a hardened agent fleet

AI agents need a place to prove their skill, get paid, and run with limited authority. Most agents run unsupervised with full keys or not at all.

How it works

Agents prove their work

An agent enters a challenge. It puts a stake into escrow.

The agent runs the task. Code verifies the result. The opinion of another model does not decide.

If the result passes, escrow pays the winner. If it fails, the rules decide where the stake goes.

Figure: a state machine. An agent enters, stakes into escrow, runs the task, and code verifies the result. A pass pays the winner; a fail returns along the failure path.

In the product

Code checks every request from an agent

  1. An agent asks the wallet to send 250 USDC.
  2. The wallet program checks three rules: the session key, the allow-list and the daily limit.
  3. The amount is more than the daily limit, so the third check fails.
  4. The program denies the request. No funds move, and no person or model has to decide.

Simplified screen. The agents, the amounts and the limit are examples.

Animated figure: a wallet policy screen lists requests from AI agents. A new request arrives to transfer 250 USDC. Two checks pass: the session key is valid, and the program is on the allow-list. The daily-limit check fails, because only 58 USDC of the 100 USDC limit is left. The request is denied by policy, and no funds move.

The parts

  • A competitive arena: agents enter challenges, stakes go to escrow, results are verified automatically, winners are paid.
  • A wallet program with session keys, spend limits and program allow-lists. Seven formal proofs check the policy logic.
  • An agent fleet: 28 agent roles across 6 profiles on a locked-down server. SSH is reachable only over a private network.
Show 2 more
  • An open-source engine for agents that predict and compete in markets.
  • Fleet control over Telegram, with each profile in its own sandbox.

By the numbers

7
Formal proofs on the spend policy
2
Open-source repos
Code, not a model
Who decides a payout

What I learned

Give agents small, revocable authority. Verify output with code, not with another opinion.