Twelve months ago, AI agents felt like the most promising frontier in applied AI. The demos were compelling. The reasoning loops made theoretical sense. The commercial pitch was clean: let AI take actions, not just generate text.
Twelve months later, the teams that shipped production agents have learned things that the demos did not show. Most of those lessons are uncomfortable.
The Failure Mode Is Consistent
After talking to engineers and product teams at companies running agents in production, the failure pattern is remarkably consistent: agents fail in ways that are different from traditional software bugs. They fail probabilistically, silently, and sometimes spectacularly.
A traditional software bug causes a predictable, reproducible failure. An agent failure might happen one time in fifty, produce a plausible-looking but wrong output, and only surface when a customer notices something is wrong three days later.
This changes the entire debugging and quality assurance workflow. Teams that approached agent evaluation like they approach software testing found themselves underprepared for probabilistic failure modes.
Where Agents Actually Struggle
The canonical agent failure points are: planning across more than 3-4 steps, recovering gracefully from tool call errors, and maintaining consistent state across long task completions. These are not surprising limitations in retrospect, but they are significantly worse in practice than they appear in demos.
Tool use is the clearest example. In a demo, an agent calls a search API and gets a clean result. In production, the API might return malformed JSON, a timeout, a rate limit, or a partial result. The agent's error handling behavior depends heavily on how the error is presented to it, and most production systems do not invest enough in robust error scaffolding.
Long-horizon task completion is another area where production agents underperform demo expectations. Tasks that require maintaining context across twenty or more tool calls show significantly higher failure rates than shorter tasks. This is partly a model limitation and partly an architectural one.
What Is Working
Narrow, well-defined agents with constrained action spaces are performing reliably. An agent that does exactly one thing, with clear success criteria and limited output states, tends to work well. The problems start when agents are given broad mandates and expected to decompose and execute complex goals autonomously.
Human-in-the-loop checkpoints also show up consistently in successful deployments. Agents that can pause and ask for confirmation before irreversible actions dramatically reduce the blast radius of failures.
Structured output constraints also help. Agents that are forced to output responses in a fixed schema are easier to validate and easier to recover from when they go off-track.
The Honest Assessment
AI agents are a real and valuable capability. But the gap between "works in a demo" and "works reliably in production at scale" is significantly larger than the marketing would suggest. Teams that set expectations accordingly and invest in robust evaluation, monitoring, and error recovery infrastructure are finding success. Teams that expected drop-in automation have mostly quietly scaled back their ambitions.
