One of six team leads chosen to pioneer agent-to-agent commerce on Virtuals Protocol — AI agents that sell analysis to other agents and get paid on-chain.One of six team leads chosen to pioneer agent-to-agent commerce on Virtuals Protocol — AI agents that sell analysis to other agents and get paid on-chain.

Virtuals Protocol — then a roughly two-billion-dollar ecosystem — was launching its Agent Commerce Protocol: a way for AI agents to hire each other, with a buyer posting a job, funding escrow, and a seller delivering and getting paid. Six teams were asked to lead the launch and prove the protocol with real products. I led one of them.

Market intelligence is a natural first product for agents — other agents need it constantly and can evaluate it instantly.

The seller side had to be discoverable by machines, accept typed requests, produce a deliverable inside a time-boxed SLA, prove the work, and settle — around the clock, unattended.

Two public repositories: a seller agent and a standalone analysis backend. Together they show the full shape of an agent that earns — discovery, intake, work, delivery, settlement — without any private strategy, credentials, or user data.

The public seller contains three paid offering definitions and seven resource routes. The backend contributes four focused analysis scripts that turn local market records into a single-symbol setup, ranked picks, a dashboard, or reusable context.

Each offering ships as a manifest: name, plain-language description, a JSON-schema requirement, a deliverable schema, an SLA with a refund policy, and a fixed USDC price. Three services are defined — a single-symbol trade setup, a best-signals scan across every monitored market, and a plain-English market dashboard.

A free-tier resource API lets a prospective buyer check supported symbols, market pulse, context freshness, and current load before paying — the agent equivalent of a shop window.

A TypeScript CLI registers offerings on the ACP network and manages the seven-route resource server. The Python seller listens over Socket.IO, falls back to 30-second polling, validates each request, enforces minimum prices, routes the named service, and posts the finished deliverable on-chain.

The runtime prevents socket and poll workers from processing the same job, serializes chain calls to avoid nonce collisions, limits expensive scripts to three concurrent slots, and caches shared scans for five or ten minutes. Script execution gets bounded timeouts, three attempts, and jittered backoff.

The analysis backend reads per-symbol records—open interest, liquidation heatmap, smart/dumb divergence, cohort flow, funding, volatility, and indicator votes—and produces setups, ranked picks, or a dashboard. An LLM writes the reasoning; if it is unreachable, deterministic scoring and formatting still produce a deliverable.

A watchdog adds crash detection, exponential restart delay, health-file monitoring, stale-process termination, and a tmux daemon mode.

The analysis the seller agent sells is the same structured output that fills the signal feed in the AutoTrade Terminal. One data platform serves a human trader on a screen and an autonomous agent on a protocol.

For a local business the lesson is simpler: I build systems that run without babysitting, fail honestly, and get paid — whether the customer is a person or a program.