Core concepts
Execution model
One request in, one report out: how routing, agents, and the pipeline turn text into findings.
An execution moves through five stages, and every stage reports whether it succeeded, how long it took, and what it produced.
- input
- The request is classified and a target is extracted.
- agents
- The agents that can serve that target are selected.
- engine
- Agents run — independent work concurrently, dependent work after.
- evidence
- Findings are collected, deduplicated, and scored.
- report
- The result is assembled, and a summary is written if configured.
Routing
You can name a target directly — /api/v1/token/0x… — or send free-form text to /api/v1/execute and let the router decide. An address that is a contract routes differently from one that is a wallet, which is why the router reads the chain before committing.
If the router cannot identify a target, the run fails cleanly with a reason. It does not guess.
Failure is partial
A failing step does not fail the task. One source being unreachable costs you that source's findings and nothing else — the rest of the report still arrives, and the missing part is named rather than silently absent.
Caching
An identical analysis inside ten minutes returns the stored result. Two identical requests arriving at the same time share one execution rather than doing the work twice.