Learning/AI Agents — Complete Guide/Lesson 28

Chapter 10·Lesson 1 of 3·10 min

Production deployment

Move an agent from prototype to reliable production software with budgets, retries, approvals, and rollback.

Concept diagram

flowchart LR
D[Candidate] --> E[Evaluation gate]
E --> S[Small rollout]
S --> O[Observe]
O --> Q{Healthy?}
Q -->|Yes| P[Expand rollout]
Q -->|No| R[Rollback]
P --> O
A[Auth] --> S
B[Budgets] --> S
C[Approval] --> S

Lesson overview

Production deployment

A demo can succeed once. A production agent must remain bounded, observable and recoverable across thousands of varied runs.

Define the contract

Document supported tasks, unsupported tasks, allowed tools, data boundaries, maximum steps, timeouts, fallback behavior and escalation rules.

Reliability

Use deadlines, bounded retries, idempotency for writes, circuit breakers and durable state for long-running tasks. Never blindly retry an operation that might have already changed state.

Cost controls

Set per-run token, time and monetary budgets. Track cost per successful task. When the budget is exhausted, stop safely or use a deterministic fallback.

Human approval

Require approval for irreversible or high-impact actions such as financial transfers, deletion, external publication and sensitive account changes.

Start with a narrow capability set. Test new tools internally, release gradually, use feature flags and maintain rollback paths. A bad prompt or tool version should be disableable without a full emergency rewrite.

Readiness checklist

Before launch verify authentication, tenant isolation, tool schemas, authorization, injection defenses, rate limits, secret handling, trace redaction, evaluation baseline, alerts, incident response and deletion/retention rules.

The goal is not maximum autonomy. The goal is predictable task completion inside explicit boundaries.

Learning path

Theory → Example → Code → Practice → Quiz → Challenge → Completion

0/6 done

Step 1

Theory

Production deployment

A demo can succeed once. A production agent must remain bounded, observable and recoverable across thousands of varied runs.

Define the contract

Document supported tasks, unsupported tasks, allowed tools, data boundaries, maximum steps, timeouts, fallback behavior and escalation rules.

Reliability

Use deadlines, bounded retries, idempotency for writes, circuit breakers and durable state for long-running tasks. Never blindly retry an operation that might have already changed state.

Cost controls

Set per-run token, time and monetary budgets. Track cost per successful task. When the budget is exhausted, stop safely or use a deterministic fallback.

Human approval

Require approval for irreversible or high-impact actions such as financial transfers, deletion, external publication and sensitive account changes.

Deployment strategy

Readiness checklist

The goal is not maximum autonomy. The goal is predictable task completion inside explicit boundaries.

Step 2

Example

Example staged release

Version 1 supports documentation search and ticket drafting. Version 2 adds ticket creation behind approval. Version 3 adds read-only CRM access after evaluation proves tenant isolation. Each capability is introduced separately so failures can be attributed to a specific change.

Step 3

Code

Run budget

typescript

const budget = { maxSteps: 8, maxToolCalls: 6, maxCostUsd: 0.05 };

function canContinue(state: State) {
  return state.steps < budget.maxSteps &&
    state.toolCalls < budget.maxToolCalls &&
    state.estimatedCostUsd < budget.maxCostUsd;
}

Budget checks are enforced by the runtime, not requested from the model.

Step 4

Practice

Write a production launch checklist covering identity, data access, tools, retries, idempotency, budgets, approval, evaluation, observability, deployment and rollback.

Step 5

Quiz

1. What should happen when the run budget is exceeded?

2. Why is idempotency important?

Step 6

Challenge

Design a staged rollout from internal users to 5%, 25% and 100%. Define the metrics and safety conditions that would stop the rollout at every stage.

Complete every stage

Work through every step in order, then the lesson will be marked complete.

Previous lessonCost and latency Next lessonRetries and idempotency

Each chapter and subtopic has its own public URL under /ai-agent.