Back to writing

AI Agents

Building Production AI Agents

A practical guide to moving from agent demos to systems with state, permissions, approvals, retries, and auditability.

8 min read29 June 2026Ubaith Sherif

Why agent demos fail in production

A useful agent is not only a prompt with tools. It needs durable state, a clear execution contract, access control, retries, and a way to explain what happened after the fact.

The moment an agent touches business workflows, the product needs approval checkpoints and traceability.

A production-friendly agent shape

A strong baseline separates orchestration, tools, policy, and user-visible audit logs.

type AgentRun = {
  id: string;
  workflowId: string;
  status: "queued" | "running" | "awaiting_approval" | "completed" | "failed";
  toolCalls: ToolCall[];
  auditEvents: AuditEvent[];
};

Approval loops

Human-in-the-loop design is a control surface for deciding where the agent can act and where business judgment is required.

Request
Plan
Tool call
Approval
Audit log
Action

Key takeaways

  • State and auditability matter as much as model capability.
  • Approval gates make autonomy safer and more useful.
  • Agent products need operational dashboards, not only chat interfaces.

Related articles

LangGraph in Production

7 min read

Designing AI Automation Platforms

8 min read