Docs · The doctrine · Section 09
Execution resilience
An agentic system composes nondeterministic, fallible dependencies: model, sidecars, external services. Resilience is designed in by default, not after the incident.
Classify the error before reacting
Each failure is first classified, because the right response depends on the type of error. We distinguish transient errors, validation errors, and business errors.
| Error type | Strategy |
|---|---|
| Transient (network, timeout, overload) | Retry with bounded exponential backoff. |
| Validation (output not conforming to the contract) | Single retry with the diagnostic reinjected into the context. |
| Business (missing resource, impossible action) | Delegation to a diagnostic step, or escalation for human review. |
| Model provider unavailable | Automatic failover to a fallback provider, behind the same port. |
| Non-critical service unavailable | Degraded mode; the application continues without the affected function. |
1 Fail-open for the non-critical, fail-closed for the sensitive
A non-critical dependency that goes down must not bring down the system: cache, observability, enrichment switch to silent degraded mode. Conversely, a sensitive action (mutation, write, external push) fails closed and explicitly rather than executing in doubt. The rule: degrade reads, never actions.
2 Feature detection and bimodal deployment
At startup, the system detects the services actually available according to the configuration. A missing dependency cleanly disables the associated function, without crashing. The result is a bimodal deployment: a minimal mode that runs on the bare essentials, and a full mode when all dependencies are present. The interface signals what is active rather than failing silently.