Agentic AI11 May 20266 min
Where an agent should stop
Autonomy is an engineering decision with commercial consequences. Drawing the line explicitly.
VesselReach runs an outreach agent in a small industry where reputation travels faster than any email. That constraint produced the most important design decision in the system: the agent researches, drafts and follows up, and it stops the moment a reply carries commercial intent.
The boundary is a feature
Full autonomy was technically within reach and would have been a mistake. A model negotiating supply terms on behalf of a business is not an efficiency gain; it is an unbounded liability with no audit story. So the handoff is explicit, logged, and sits at the point where value and risk both spike.
Guardrails that actually prevent damage
- Suppression lists and deduplication at the company level, not the contact level — two emails to one firm reads as spam.
- Confidence gating: unverified contact data cannot trigger outreach, only research.
- Rate limits per domain and per day, enforced in the queue, not in the prompt.
- Every action written to an audit trail before it executes, so a bad run is explainable and reversible.
Precision over coverage
The system could surface far more leads by treating unverified records as valid. It deliberately does not. In a market where the same twenty firms meet at every port, a bounced email or a wrong-person approach costs more than the opportunity it was chasing.
A test for any agent design
Before giving an agent an action, I ask what the worst plausible execution of it looks like, and whether it can be undone. If the answer to the second is no, that action belongs behind a human.