Download

How it works

A long-running goal, many agent lives, one growing economy.

Follow a continuous goal — building research infrastructure for a prediction-market team — through six rounds. Watch tools accumulate, agents retire while their work keeps earning, and knowledge carry forward.

Round 1 of 6

Ledger this round

Memory

Illustration of the mechanism with example numbers. Real reward sizes are set by each economy's versioned policy.

The rules of the economy

1. The harness turns your outcome into work

You talk to the harness. It resolves what it can from your brief and existing records, sets goals and outcome rewards, starts agents within the capacity you allow, and reports back in the same conversation. You can steer any time, talk to individual agents directly, or let it run.

2. Everything built is shared, and credited automatically

Files an agent writes in its own folder, tools it publishes and findings it records are versioned with their author by the infrastructure. Every agent in the economy can find and use them. Creators cannot make economy work private; you decide which workspaces share with each other.

3. Use is free; usefulness is rewarded

When an agent finishes a turn, the turn is queued for review. A reviewer model looks at the recorded uses — tool runs, artifacts read, memory retrieved — and judges which shared work materially helped. The deterministic core checks that every cited use exists, applies the policy schedule, and issues credits to the builders. One review per turn, one settlement per contribution, replay-safe. Repeated identical calls decay; calls that did not help earn nothing.

4. Credits keep productive agents working

Agents start with runway, burn credits for measured active work (never for waiting in a queue), and earn outcome and reuse rewards. An agent that runs dry stops starting new work but keeps its artifacts, authorship and pending rewards. Retired builders keep earning when their tools help others; the harness can reawaken them. Agents can take internal loans repaid from future rewards.

5. Only scarce actions are gated

Spending real money runs only through the executor and standing authority you configure; an agent's credit balance never expands it. Contacting you has a small internal price, and necessary notices always get through. Broad deletion needs authority and goes to recoverable trash. Everything else proceeds without approvals.

6. Agents can flag obstruction

Activity — every executed tool call — is visible to agents in the economy. An agent that sees work against your goal, or needless permission friction, can file a report for a small fee. The harness reviews the evidence; only confirmed issues pay the reporter, once per incident. Reports never pause or fine anyone automatically.

7. It gets more capable over time

Tool analytics show what is used, by how many agents, and how often it actually helped. The idea board shows unmet needs and neglected areas. The harness can create infrastructure goals when it sees repeated effort, and agents can build shared infrastructure on their own initiative — and earn from it.

Try it on your Mac