Production patterns
Moving an agent from demo to system means making execution limits, side effects, recovery, isolation, and quality controls explicit. Agent RT provides these as runtime primitives.
Bound every run
Define turn, tool-call, elapsed-time, and token budgets. Cancellation should propagate through model calls, streaming, and tool execution rather than being simulated in application code.
Make side effects resumable
For long-running work, pair checkpoints with an append-only event history and stable idempotency keys. Retries and resumed runs can then reuse earlier side-effect results instead of duplicating them.
Execute code behind a sandbox
Command and code execution belongs behind an injected sandbox boundary with explicit workspace, resource, and network policy. Agent RT supports restricted native execution, Docker, and optional managed/remote backends; unsupported isolation requirements should fail closed.
Observe and evaluate
Production operations include traces, logs, metrics, usage/cost accounting, deterministic graders, model judges, tool-use and trajectory grading, safety suites, golden datasets, and controlled experiments.
Separate deployment from agent logic
Deployment and optimization primitives cover configuration overlays, feature flags, prompt rollout/rollback, workers, quotas, schema migrations, caching, batching, latency optimization, and multiple deployment targets. These concerns can evolve without changing the agent's provider-neutral core contracts.