The gap between an agent that demos well and one you would put in production is not smarter prompts — it is operations. Real agents need to be orchestrated, constrained, watched, measured, and made to fail gracefully. Skip this and you ship something that works on the happy path and quietly does damage on the others.

This is the discipline that turns a clever loop into a dependable system.

Five load-bearing practices:

- Orchestration patterns — how work is coordinated: a single loop, prompt-chaining, routing/classification, parallelization, or orchestrator-worker (supervisor delegating to workers, from AIS-02). - Human-in-the-loop (HITL) — insert approval gates before high-stakes or irreversible actions; the agent proposes, a human confirms. - Guardrails — constraints on inputs and outputs: content filters, allowed-tool lists, output schemas, and checks that catch policy violations or unsafe actions before they execute. - Evaluation (evals) — systematic, repeatable tests of agent behavior on a curated dataset, scoring task success, not just vibes. This is how you know a change helped or hurt. - Observability / tracing — capturing every step, tool call, token, latency, and cost as a trace, so you can debug and monitor what the agent actually did.

Failure handling ties them together: timeouts, retries with backoff, fallbacks, and graceful degradation when a tool or step fails.

A support agent can issue refunds. Production hardening:

- Guardrail: an input filter rejects prompt-injection and out-of-scope requests; an output schema forces the refund amount into a validated field. - HITL: any refund over 100 dollars pauses for a human to approve. - Tracing: every run logs the reasoning steps, the tools called, latency, and token cost. - Eval: a dataset of 200 past tickets scores whether the agent chose the right action; you re-run it on every prompt change to catch regressions. - Failure handling: if the payments API times out, the agent retries twice with backoff, then falls back to creating a human ticket instead of guessing.

Same agent, but now observable, bounded, tested, and safe to fail.

These layers wrap everything else in the path. Orchestration coordinates the agents from AIS-02; guardrails and HITL make the tool use of AIS-03 safe; evals and tracing measure whether RAG grounding (AIS-05) actually reduced errors.

The throughline: you cannot improve what you cannot see, and you cannot ship what you cannot bound. Tracing gives you visibility, evals give you a feedback signal, guardrails and HITL give you bounds, and failure handling keeps the system standing when a dependency misbehaves.