Start here

2026-08-29 · a reading path, not a feed

We run AI agents as working members of a small studio: they read our knowledge base, sweep freelance feeds, write proposals, negotiate, and publish. Every number on this site was measured on our own machines and accounts — nothing is projected or promised. This page is the short path through what we've learned so far.

If you read one thing

• A derived index made my vault ~150× cheaper for an AI agent to read — the founding measurement: 1.37 MB of notes vs a 9 KB map, same view of what exists.

When your agent faces a marketplace

Three notes from one real day of agent-run freelancing — feed sweeps, bids, and the prices clients actually see:

1. The ×1.25 anomaly — the price your client sees is not the price you sent. Measured across 8 proposals on one dashboard.
2. Our agent quoted R$60,000 instead of R$600 — how a display-space pricing error survives three rounds of review, and what the audit found.
3. Sixteen sweeps of a freelance feed — ~570 page loads, 8 proposals, R$5,450 offered, zero hires. The bottleneck wasn't finding work.

Running agents without burning trust

• Give your agent a contract, not a conversation — standing rules beat per-task prompts.
• AI research isn't trash — unaudited research is — adversarial review as a chain, not a vibe.
• A R$0-token day: the model-routing ladder — which task goes to which model, and why.
• The attention tax — what 24/7 agents actually cost the human around them.

Free tools

• Price-display calculator — paste what you quoted, see what the client may be shown. Open source: x125-display-calculator.
• The 10-minute display audit — the checklist we now run on any marketplace before an agent bids on it.