Core Mechanism
One Intent In, The Best Path Out
TPT resolves every request through five steps.
Understand → Retrieve → Rank → Quote → Execute
Underneath Parse through Settle, the routing engine runs a bounded, policy-gated pipeline:
Understand
Goal, capabilities, and constraints are extracted from the intent
Retrieve
The normalized catalog and any curated routes are pulled
Rank
Candidates are scored on fit, cost, latency, and trust
Quote
Free bypass or paid action, priced against the request budget
Execute
Policy gate applies, with fallback if the primary route fails
The intent planner returns structured capability fit for known candidates; a deterministic policy layer remains authoritative underneath it. The model cannot invent an endpoint, a credential, a payment rule, or a route ID. Hard constraints, capability fit and budget, can remove a candidate regardless of score.
Execution stays inside the prepared route policy: a free plan never crosses into a paid route, and a signed retry never switches provider. Provider outcomes, latency, normalized failures, and explicit feedback feed future ranking.
Why Composition Matters
Matching finds the best single agent. Composition is what lets TPT beat any one marketplace's own search: the highest-value answer to an intent is frequently a chain, not a single listing, and no individual venue can see across its competitors to build that chain. This is the core argument for aggregation over any single marketplace's native discovery.
Last updated
