How an AI agent discovers, reads, and hands off to Spectare

Agents interact with Spectare in three sequential steps. First, discovery: your site serves an llms.txt file that describes what your organisation offers, which audiences you serve, and which endpoint to call. An agent that finds this file knows how to qualify a visitor without any custom integration work.

Second, qualification: the agent takes what the user actually said and posts it as a plain-text summary to POST /api/qualify/{orgSlug}. Spectare classifies the intent independently using Claude, then returns a signed context token bound to that workspace. This is the step that matters for cache behaviour. A token produced by /api/qualify carries a proven org claim, so it can write to the shared cache other visitors read and count toward the learning loop. Agents should send the user's own words here rather than a pre-formed classification: the point of the qualify step is that Spectare makes the judgement, not the agent.

Third, assembly and handoff: the agent passes the token to GET /api/assemble/{orgSlug} or constructs a landing URL using the token as a query parameter. The assembled page is built from that workspace's real atoms, personalised to the classified intent, with no further input needed from the agent. If Spectare's classification differs from what the agent inferred, the page reflects Spectare's version, which is the property that keeps agent-sourced content safe to serve to real visitors.