Each of these is a sequence an agent can run on its own. The first three are also built in as prompts, so a client can offer them as commands.
Find something tradeable, then prove it
Four to ten API requests in total. Report the pair, what it costs, what it pays, the fees, the size the books showed, and what would kill it. depth.limited_by names what stopped the fill growing: price, depth, minimum, fees or unavailable.
Price one question on both venues
Two requests. compare_venues matches outcomes across venues by proven equivalence, not by wording, so “Fed cuts 25bps” on one venue lines up with its counterpart on the other. A gap is not a trade: check find_opportunities for whether it can be filled after fees.
Watch one market and say when something happens
Signals describe trading, not its cause: burst is far more trades than the market normally sees, large_fill is one fill much larger than its typical trade, one_sided is unusual flow almost all one way that moved the price with it. Say what moved and how much money was behind it; do not invent a reason.
For a standing watch, do not poll every minute — subscribe to the lead.executable webhook and let the agent react to the delivery.
Test an idea against settled history
Calibration is the honest version: for each 10-cent band, how many settled markets fell in it, what they averaged, and the share that resolved Yes. A band whose yes_rate sits above its avg_price_usd is where a strategy’s return comes from. Buying Yes at price p in a band that resolves Yes at rate r returns 100 * (r / p - 1) per $100 before fees.
Returns are historical, before fees, and use each band’s average midpoint rather than the ask you would have paid. A rule fitted on the history that selected it is not an out-of-sample test, and a 95% range is not a floor.
Writing the agent’s instructions
Give the agent these rules; they prevent most wrong answers.
Handling limits
A tool call that runs into your key’s limit comes back as a tool error naming rate_limited or daily_quota_exceeded with the seconds to wait. Teach the agent to wait rather than retry immediately, and remember that one composed call can cost several requests — each tool’s description says how many. See Rate limits. 
