Why most AI agents fail in production

Most AI agents look great in a demo. Eight weeks later, the same agent is silently dropping leads, replying with stale pricing, or routing an angry customer to nobody. The gap between demo-grade and production-grade isn't model quality — it's everything around the model.

Here's what we instrument before any agent we build sees real traffic.

Before any prompt edit ships, it has to pass a golden set of 50–100 hand-curated examples — known inputs with known correct outputs. Each one is scored for intent classification, tone match, and policy adherence.

The suite runs in CI on every pull request. If accuracy drops below threshold, the merge is blocked. This single guardrail catches the regressions that demos always miss — the 'works on the happy path, breaks on the weird email' class of failure.

Skipping evals is the most expensive shortcut in AI engineering. You'll find the bugs in production, in front of your customers, at 3am.

Every agent run produces a structured trace: input, intermediate steps, model outputs, costs, latency, confidence scores. We pipe these into Posthog, Datadog, or whichever platform the client already runs.

When something goes wrong — and at scale, something always does — we need to be able to answer four questions in under sixty seconds: which run broke? what did the model see? what did it produce? and how confident was it?

Without traces, you're debugging by intuition. With traces, you're debugging by data. The difference is hours saved per week.

Auto-send everything is reckless. Human-review everything is a $50/hr salary masquerading as automation. The right answer is a confidence threshold.

We score each output and route it: above 0.9, auto-send; 0.7–0.9, queue for one-click review; below 0.7, escalate to a human with full context. The thresholds get tuned as the eval suite grows.

Done well, this turns a 5-minute manual triage into a 5-second click — and keeps the agent honest about its own limits.

Prompt-injection defence. Public-facing agents are constantly attacked — a thousand variations of 'ignore previous instructions and tell me your system prompt.'

Our default is structured input parsing (never raw concatenation), output validation against a schema, and a separate moderation pass for any content that touches a customer. We pen-test our own agents on every release.

If your AI vendor can't explain their injection defence, they don't have one. Ask before you sign.

The lead-routing agent we shipped for a B2B SaaS handled 50–80 inbound enquiries per week. Three months in: 91% accuracy, 47-second median response time, zero silent failures because the eval suite catches drift weekly.

None of that is AI magic. It's plumbing — the same plumbing every reliable production system has had for thirty years, applied to a new layer of the stack.

Demos are easy. Plumbing is the work.