Copy this template for every design.
1. Outcome
User:
Job to be done:
Business outcome:
Success metric:
Non-goals:
2. Why agentic?
Why a deterministic workflow is insufficient:
Where model judgment is useful:
Where deterministic code remains authoritative:
Simpler alternative considered:
3. Risk and authority
Risk tier:
Worst credible failure:
Read-only actions:
Reversible writes:
Irreversible/material actions:
Human approval points:
Prohibited autonomous decisions:
4. Journey and topology
Happy path:
Agent/workflow pattern:
Key states:
Stop condition:
Escalation condition:
Cancellation behavior:
5. Tools
| Tool | Input/output contract | Identity/scope | Side effect | Idempotency | Timeout/retry | Approval |
|---|---|---|---|---|---|---|
6. Context and data
Trusted instructions:
User inputs:
Retrieved sources:
Authority/freshness ranking:
ACL strategy:
Citation/provenance:
Context budget/compression:
Untrusted-content boundary:
7. State, memory, and durability
Request state:
Run state:
Thread/session state:
Workflow history:
Long-term memory:
System of record:
What survives a crash:
What must never repeat:
Resume/version strategy:
8. Security and governance
Trust boundaries:
Tenant isolation:
Secrets and service identity:
Prompt-injection controls:
Data-exfiltration controls:
Sandbox/egress controls:
Retention/deletion:
Audit events:
9. Evals
Golden cases:
Adversarial cases:
Component metrics:
Trajectory metrics:
Outcome metrics:
Safety/policy metrics:
Human calibration:
Release gate:
10. Operations
Latency SLO:
Availability/durability SLO:
Cost budget:
Throughput/peak load:
Rate-limit/backpressure strategy:
Critical alerts:
Runbook owner:
11. Rollout
Offline test:
Shadow:
Canary:
Rollback trigger:
Versions recorded:
Production sampling:
12. Open tradeoffs
Interview shorthand
Under time pressure, cover these in order:
- outcome and risk;
- authority boundaries;
- topology;
- tools and context;
- state and durable execution;
- security and human review;
- evals and SLOs;
- scale, cost, and rollout.