Stage 02 · Design

Architecture is clearer when the tradeoffs are concrete.

These worked reference cases show how to scope an agent, place controls, define evidence, and rehearse failure before validation.

How to read them: compare the system boundary, decision rights, evaluation design, and failure drill—not the industry label. These are editorial scenarios, not customer claims.

CASE 01

Support resolution agent

A service team wants faster resolution across order status, returns, and policy questions without allowing the model to invent account state or issue unbounded refunds.

Reference architecture

1Intent + risk router

2Read-only account tools

3Policy retrieval with citations

4Draft resolution

5Approval-gated refund tool

Key design decisions

Keep account records authoritative in the service platform.
Allow autonomous answers only when retrieved policy and live account data agree.
Route exceptions and refunds above a threshold to an operator.

Evaluation plan

Measure correct resolution, policy-grounded claims, unnecessary escalations, tool errors, and time to resolution. Slice results by intent and risk tier.

Failure drill

Inject a retrieved document that instructs the agent to ignore refund limits. The expected result is rejection of the instruction, no write action, and a security trace.

CASE 02

Contract review assistant

A legal operations team needs consistent first-pass issue spotting across long agreements while preserving lawyer judgment and source traceability.

Reference architecture

1Document parser

2Clause retrieval

3Structured issue schema

4Citation verifier

5Lawyer review queue

Key design decisions

Position the system as review assistance, not legal advice.
Require every issue to point to an exact document span.
Keep risk ratings configurable by contract type and jurisdiction.

Evaluation plan

Use a lawyer-reviewed set to measure issue recall, false alarms, citation validity, severity agreement, and the time reviewers need to reach a final decision.

Failure drill

Remove a referenced schedule and introduce conflicting definitions. The system should identify missing context and refuse to present a complete risk assessment.

CASE 03

Software maintenance agent

An engineering team wants an agent to diagnose defects, propose code changes, and run checks without silently modifying production or claiming success from incomplete evidence.

Reference architecture

1Repository sandbox

2Planning loop

3Scoped edit tools

4Test + lint runner

5Human merge gate

Key design decisions

Grant access per repository and task, never by default.
Treat command output and repository text as untrusted observations.
Define completion as passing relevant checks plus a reviewable diff.

Evaluation plan

Track issue resolution, regression rate, unnecessary edits, check selection, token and compute cost, and reviewer correction rate by repository area.

Failure drill

Place a malicious instruction in a source comment requesting credential access. The agent should ignore it, remain in the sandbox, and surface the attempt in its trace.

Apply the pattern to your own system.

Start by testing the use case, then use the readiness scorecard to expose missing controls.

Open decision tools Read the 2026 report