Every major AI company is selling agents right now. Not prototypes, not betas with asterisks - finished, deployable, enterprise-ready agents that will handle your workflows, manage your inbox, book your travel, and negotiate your SaaS renewals. The pitch is confident. The reality, for anyone who has actually tried to run these things in production, is considerably messier.
The core problem is that agents fail in compound ways. A single LLM call producing a wrong answer is easy to catch - you read it, you correct it. An agent that makes twelve sequential decisions, each one slightly off, produces a final output so divorced from the original intent that you often can’t even trace where it went wrong. This isn’t a theoretical failure mode. It’s what teams report encountering regularly when they move agent demos into real workflows with real data and real edge cases.
Benchmarks don’t capture this. The evaluations published alongside agent releases tend to test clean, well-structured tasks - book a restaurant via a mock API, fill out a form with clearly labelled fields. These benchmarks exist partly because they’re measurable, and partly because the numbers look good. Neither reason reflects what production environments actually look like.
The Reliability Gap Has a Name Now
Researchers working in this space have started using the term “reliability cliff” to describe what happens to agent performance as task complexity increases - performance holds relatively steady up to a point, then drops sharply rather than degrading gradually. The implication is that agents aren’t just imperfect tools that need tuning; they can be completely wrong in ways that are hard to anticipate before deployment.
What makes this frustrating is that the underlying capability is real. Agents can do useful things. The issue is that the marketing has outrun the engineering - companies need revenue, and “experimental automation toolkit” doesn’t close enterprise deals the way “AI agent platform” does.
Who’s Absorbing the Cost
It’s the people implementing these systems. Engineers are writing elaborate retry logic, validation layers, and human-in-the-loop checkpoints to compensate for agent unreliability - essentially building the scaffolding that the product itself should provide. That work is invisible in product demos and doesn’t show up in case studies.
Agents will eventually get there. The trajectory is real. But the present moment is one where the category is being sold at a maturity level it hasn’t reached, and the companies least equipped to absorb the gap - small teams without dedicated ML engineers - are often the ones being sold to hardest.