Set the intent
Define the offer, audience, voice, proof points, prohibited claims, and approval rules. This is the human context the agent should not have to rediscover.
memory → audience → boundariesHow it works / the handoff
Intent in.
Clarity out.
You set the intent once. The agent carries research, preparation, readiness, and reply handling forward; you only enter where the product asks for judgment.
See the agent contract ↗CONCEPTUAL WORKFLOW
brief→research→preview→review
Illustrative sequence · not live runtime evidenceAgent carries
Research, prepare, recover.
It gathers bounded evidence, assembles a preview, checks readiness, and reports what it could not verify.
Inspect the contractHuman decides
Intent, exceptions, approval.
You define the boundary and decide when a prepared action is allowed to cross it.
See the operating layer01 / The path
One thread,
fewer blind spots.
Research with a boundary
Request a bounded lead or account pack. The result carries provenance, freshness, confidence, suppression eligibility, and a review state.
source → evidence → candidate statePrepare, then preview
Draft a sequence from approved context. Validate variables and risky claims. The preview shows affected leads, warnings, and an evidence identifier.
draft → validate → dry-run receiptApprove the edge
Readiness, limits, suppression, and human approval determine whether an action may move forward. If the gate fails, the system returns a typed reason.
policy → approval → apply or holdTurn replies into next actions
Ingest a reply, classify it, stop a sequence where appropriate, and create a next action with enough context for a person to decide quickly.
reply → classification → next actionThe point
“The agent did the middle. The team still knows what happened.”
A design principle for delegation, not a customer testimonial.
Start with an intent, not a form