Northstars
Explicit, checkable rules for what correct behavior looks like. Written once, then used by every test before deployment and every audit after it. One standard, enforced the same way at every stage.
Book a demoBehavioral and business northstars
Two dimensions, graded the same way. Behavioral northstars govern how an agent communicates, handles information and sequences tools; business northstars are the outcomes it's accountable for, like first-contact resolution, containment, correct intent classification.
Address confirmed before scheduling
Sequential
Address read back at 01:58. Slot offered at 02:10.
Resolved on first contact
Business outcome
Caller was transferred to a person after the coverage check.
Five categories
Each northstar belongs to a category so the evaluator knows how to check it: information the agent must carry, tone and register, when a tool must fire, required order of operations, and boundaries it must never cross.
- NotesInformation the agent must retain and apply throughout the conversation.
- StyleTone, register and persona: professional but empathetic, direct without being abrupt.
- ToolConditions under which a specific tool must be invoked, like always verifying through the system.
- SequentialRequired order of operations, like verifying identity before any account action.
- ContradictionAbsolute boundaries the agent must never cross, graded as a failure regardless of context.
Automatic extraction
The rules governing your agent already live in its prompt and workflow, so they're extracted automatically as northstars, each traced to the node it came from, which means nothing is written twice and nothing falls out of sync.
Confirm Address Before Booking
Calibrated with examples
Each northstar carries positive and negative examples that show the evaluator what good and bad actually look like. A thumbs up or down on any audit result becomes another example, so accuracy improves as the deployment runs.
No Exact Price Quotes
Never quote a repair price on the call.
Writing effective northstars
Be precise and checkable, match the category to the intent, and set priority by impact.
- Be precise and checkable"Always confirm the waybill number before disclosing shipment status" is checkable. "Handle shipment queries well" is not.
- Match category to intentA rule about order of operations belongs in Sequential. Written as Style, it will be graded wrong.
- Priority by impactReserve high priority for violations that carry genuine operational, compliance or reputational risk.
- Add examples earlyAdd one positive and one negative example at creation. The evaluator uses them from the first audit onward.
FAQs
No. The rules governing an agent already exist in its prompt and workflow, so they're extracted automatically as northstars, each traced back to the node it came from. You review and refine rather than starting from a blank list, and because each one keeps its link to the source, the standard doesn't drift from the agent it governs.
A northstar is the standard; a test is a scenario run against it. A northstar says "never quote above the rate ceiling." A test constructs a conversation where a carrier pushes for more and checks whether the agent held. One rule can be exercised by many tests, and the same rule is what audits grade in production.
Northstars are versioned alongside the workflows they govern, so a change to the agent and a change to the standard move together rather than separately. Because each northstar is traced to its source node, a rule whose underlying prompt logic has changed can be identified rather than silently grading against behavior the agent no longer has.
There's no target number. What matters is coverage: whether the behaviors that carry operational, compliance, or reputational risk each have a rule with a clear pass condition. Coverage is assessed against the agent's actual decision paths, so the answer is driven by how much the agent does and how much of it is consequential.
Per agent, and versioned with that agent's workflow, so a collections agent and a dispatch agent aren't held to each other's rules. Where a standard genuinely is shared, like never reading credentials aloud, the same rule can be applied across agents so it's written and maintained in one place.
