From Demo to Production: The AI Agent Hardening Guide

Every AI agent demo works. That’s the problem.

In the demo, the data is clean, the task is scripted, and the agent looks flawless. Then it meets production: messy inputs, edge cases nobody scripted, a tool that returns something unexpected at 2 AM — and it fails silently, in ways nobody notices until the damage is done.

The distance between a demo and a production agent isn’t the model. It’s the engineering around the model. After building agents that run inside real businesses, I harden every build with the same six layers:

1. Guardrails — scoped permissions, allowlists, sandboxed tools, least-privilege keys. The agent can only touch what its job requires — nothing else exists for it.
2. Approval gates — money and irreversible actions never execute autonomously. The agent proposes, a human approves with one tap, then it acts.
3. Observability — immutable logs and a trace for every run, so “what did the agent do?” always has an exact answer.
4. Evals — a fixed test set runs before every change, against a target score. If a tweak makes the agent worse, you know before your customers do.
5. Fallbacks & recovery — retries, graceful degradation, human handoff, undo where undo exists. Failure is planned, not improvised.
6. Cost controls — per-run and per-day spend ceilings, single-use payment instruments, and a kill switch that actually kills.

Skip these and the math turns ugly fast: one unsupervised refund or purchase at the wrong amount costs more than the approval gate that would have caught it — and a silent failure doesn’t just lose an hour, it quietly compounds for days before anyone sees it.

The guide goes layer by layer with the concrete mechanics — permission scopes and deny-by-default tool registries, the one-tap approval pattern and single-use virtual cards, append-only run logs and per-run traces, a fixed eval set with a written target score and disaster cases, the failure table you write before launch, and the per-run and per-day ceilings that bound a bad day. It finishes with the full roadmap — define the job and its scope, scaffold a hardened skeleton with v0, Lovable, Bolt or Replit, wire the model APIs properly, add observability and the approval channel, build the eval set from real cases, then deploy behind tight ceilings and loosen only as the logged track record earns it — plus the 7-point checklist: scoped permissions, approval gates, immutable logs, eval set, human fallback, cost ceiling, kill switch. Download it above and ship agents that survive production.