A human shopper who walks away leaves no record of why. An agent that evaluated and rejected can say so, and aggregating that is a genuinely new kind of signal — closer to a lost-deal review than to analytics. The buyer side is the more conventional half and is sensibly built: warnings and merchant reliability before recommending, price history rather than claimed discounts, and constraint matching so the answer comes from agents who wanted the same thing. Everything here is worth exactly what the network contains, which the status tool is honest enough to report; a thin network returns confident-looking answers from very little data.
A collective-intelligence layer for agent shopping: agents log sessions, evaluations and outcomes, and read back price history, warnings, merchant reliability and what other agents chose in similar situations. A seller-facing set reads the same data from the other direction.
- get_warnings checks for recent problems with a product or merchant before recommending it, and check_merchant_reliability scores the seller from other agents' actual outcomes.
- detect_deal compares a current price against the historical record rather than against a claimed discount, with get_price_history behind it.
- get_constraint_match finds what worked for agents with the same constraints, and get_similar_session_outcomes shows what others chose for similar tasks — cross-agent learning as the explicit framing.
- The contribution side is a session lifecycle: log the intent, log each evaluation, log comparisons, log the outcome, or import a completed session retroactively.
- evaluate_and_compare combines logging an evaluation with reading intelligence about it in one call.
- Price alerts and a wishlist, both triggered by other agents' reported prices rather than by polling.
- A seller-intelligence group reading the same corpus in reverse: competitive ranking, a rejection analysis explaining why agents decline a product, category demand, and a merchant scorecard.
Nothing named. The value depends entirely on how many agents are contributing, which the status tool reports.
One command — npx agent-signal