Most AI agents aren't production-ready
52% of enterprises say they have AI agents in production. Only 5% are worried about tool-calling accuracy. The gap between deployed and reliable is where the real work happens.
The numbers in the latest Anthropic enterprise survey are striking: most organizations claim to have agents in production, but very few have measured whether those agents actually do what they're supposed to.
Here's the uncomfortable truth: shipping an agent that works on the happy path is now a few weekends of work. Shipping one that works the thousandth time, on the unhappy path, with the right escalation when the tool call fails — that's still mostly hand-written engineering.
The four levels of "deployed"
- Demo — works in the boardroom, breaks on the second customer.
- Beta — works most of the time; you have an angry Slack channel for failures.
- Production — works reliably; you have observability, retries, fallbacks, and a clear escalation path.
- Compounding — improves over time because you have an eval harness, prompt versioning, and a feedback loop.
Most "deployed" agents I see are at level 2.
What level-3+ looks like
- A formal eval set that lives next to the code, runs in CI, and blocks model upgrades that regress
- Per-tool-call telemetry: latency, cost, success/failure, downstream impact
- Constrained outputs (JSON schema or grammar) for any structured field
- A degradation path: if the agent fails three times, it hands off to a human with full context
- Cost guardrails enforced server-side
The good news: every one of these is a solved problem. The bad news: nobody is going to install them for you, and most teams underestimate how many of them they need.
Keep reading