Agent Sensor Stack
An agent should be surrounded by sensors. And crucially: the agent doesn’t get to declare success. The sensors declare success.
┌──────────────┐
│ AGENT │
│ writes code │
└──────┬───────┘
│
┌─────────────┼──────────────┐
▼ ▼ ▼
compiler tests linter
│ │ │
└─────────────┼──────────────┘
▼
mutation test
│
▼
integration test
│
▼
canary sensor
│
▼
production events
│
▼
outcome sensors
The agent sensor stack is the confidence stack applied to AI-generated code: each layer is a gate the agent’s output must pass before it reaches production. The key principle is that the agent cannot self-certify — independence must be preserved at every layer.
Sensor properties
| Property | Value |
|---|---|
| Oracle strength | High — the combined stack is strong |
| Independence | High — if producer and evaluator are separated |
| Scope | System-level |
| Feedback latency | Varies by layer |
| Actionability | Guiding — each layer tells the agent what failed |
| Type | Predictive + Retrospective |
What it cannot detect
The stack is only as strong as its weakest layer. If the agent writes the tests (low independence), the test layer provides no real signal. See producer-evaluator separation.