How agents hunt

Every enabled agent runs unattended on a schedule. Daily for most types, weekly for the market mapper, whose deep runs spend real money. A run is a fixed pipeline, and knowing its stages explains both what lands on your board and what you're paying for.

1. Fetch

The agent sweeps its corner of the internet: a job board's API, a subreddit, SEC filings, a news query. Fetching is free for almost every type; only the two Claude scouts spend model time to find candidates in the first place. Every fetched item carries a stable identity, so the same posting or company is never considered twice across runs.

2. Prefilter, free and deterministic

Before any money is spent, deterministic filters drop what can't fit: duplicates of anything this agent has already seen, companies outside your size bounds, anything matching your hard excludes. No model involved.

3. Judge

Survivors go to a Claude judge in batches, with your full ICP brief as the bar. Each candidate gets a score from 0 to 100, a one-line reason, and any disqualifiers. That reason is what you see on the card, so every lead arrives with its "why". Judging is the main model spend of a normal run, and it's bounded: fetches are capped per run, so a busy day can't blow the budget.

4. Threshold and daily cap

Leads at or above your threshold become cards, newest first, until the agent's daily cap is reached. Everything else is recorded with its verdict (below threshold, capped, duplicate), so you can audit what the agent declined.

5. Person resolution, opt-in

If enabled for the agent, accepted leads go through person resolution: a search for the actual human matching your persona, returning a name, title, and an email with a confidence score and evidence. This spends money per lead, which is why it's off by default and runs only on accepted leads, never on the whole fetch.

6. Drafting and auto-send, opt-in

An agent can go one step further and draft outreach for each accepted lead, or send it behind a separate, higher score bar. Both are off by default. See Auto-send for how the send bar works.

Budgets

All model spend (judging, resolution, drafting) draws from one monthly team budget. When it runs out, agents stop adding leads rather than degrading quietly, and every run records exactly what it cost.