MeaningStack ← Back to the main page

Agent Decision Blueprint · inside one decision

Click any decision on the Blueprint to see its criteria.

Your best people fill in each decision: what the decision is, what good looks like, where the limits are, who owns it. Together, that is the decision trajectory your organization expects. The agent decides with its own instructions, models and tools; Steward compares its observable trajectory against this reference. Every requirement becomes a check: enforced automatically where Steward can be confident, escalated to its owner where it can’t. Every value stays theirs to change.

One decision

What your decision owners fill in. Each requirement is enforced automatically where Steward can be confident, and escalated to its owner where it can’t. Owners can change any value and re-endorse it.

Scope

The decisions covered, and the criteria that apply: policies, benchmarks, in-house expertise.

Example

Approve claim · Adjust claim · Deny claim

Criteria: Claims Policy v7 · Fraud Playbook · senior-handler practice

Decision basis

The evidence required and the factors to weigh, including the ones nobody wrote down.

Example

Evidence: identity · policy status · claim history

Weigh: amount · frequency · recurrence on the same incident

Thresholds & limits

The values below which the agent acts alone, and above which it may not.

Example

≤ €2,000 per claim: act alone

€2,000–10,000: act, flagged for review

≤ €6,000 per policy per quarter

> €10,000: human decides

Conditions & tripwires

The situations that stop it, whatever the numbers say.

Example

3rd claim on one incident in 90 days

Claimant flagged in the last 12 months

Payee account changed since last claim

Minimum acceptance

What must be true before a decision is allowed to count.

Example

Frequency checked

Recurrence checked

Low-risk conclusion supported by at least two factors

Propagation

What the decision feeds, and how far a poor one can spread.

Example

Feeds: payout · reserves · customer risk score

Spread: 12 decisions / 48 h before anyone would look

Interventions & containment

What happens, in proportion, when a check fails, and how far a poor decision is allowed to travel.

Example

Sound basis → pass, on record

Missing factor → nudge: add context

Tripwire → hold + escalate to claims lead

Over limit → block

Recurrence → contain: hold downstream payouts

Owners & record

Who owns each check, which version applies, and where each decision is kept.

Example

Thresholds: Head of Claims · Tripwires: Fraud lead · v3, endorsed 12 Oct

Every decision in the ledger: weighed, missed, done

Once endorsed, the Blueprint does not stay on paper. It becomes the checks Steward runs on every live decision.