Agents in Production
Run systems that deserve trust: service objectives, Golden Signals, telemetry, Kubernetes, Grafana, fleet configuration, resilience and security boundaries.
Production is a different job
A useful prototype proves an idea. A production service makes an ongoing promise to people who depend on it, including a promise about failure.
Golden Signals for agent systems
Start with latency, traffic, errors and saturation. Then add agent-specific measures such as quality and cost without losing the SRE signals that explain the service.
Trace the whole decision
A trace should let an operator follow one request through context selection, model calls, tools, policy decisions and final outcome without exposing secrets or private content.
Kubernetes is an execution environment, not a strategy
Kubernetes can schedule and recover workloads. It cannot choose your task boundaries, retries, data controls or release policy for you.
Grafana dashboards that lead to a decision
A useful dashboard answers an operational question quickly. It does not try to display every metric your service has ever emitted.
OpAMP: operating the telemetry fleet
Open Agent Management Protocol helps manage OpenTelemetry Collector-style agents at scale. It is about operating telemetry components, not about an AI agent making decisions.
Reliability control loops
Timeouts, retries, queues, idempotency and human escalation work together. A retry without a stop condition is just a faster way to make a bad incident bigger.
Security boundaries for agentic systems
Treat model input as untrusted, tool authority as sensitive and every integration as a supply-chain decision. Security is enforced by systems, not by asking a model to behave.
Release and incident readiness
Before a model-enabled feature is trusted in production, prove that it can be evaluated, observed, constrained, rolled back and supported when it fails.