Agents in production: one year on
Of the seven agentic flows we launched in 2024, five are still alive. The two that died went the same way.
We shipped seven distinct agentic flows in 2024. Autonomous systems that planned work across multiple steps and took action without human intervention throughout their execution. A year later, five of them are still running in production and working correctly.
The two systems that failed followed the same failure pattern consistently: both depended on third-party websites not changing their HTML structure or interface. Both websites changed their structure in ways that broke the automated interaction immediately and completely.
The agents broke silently without alerting anyone in advance when the interface changed unexpectedly. The output became garbage and unusable. Eventually someone noticed the failures had accumulated and caused problems in production.
The survivors run against stable, versioned APIs from the client's own systems and include a manual retry path for failures. When they encounter unexpected responses they flag the situation for human review and intervention.
The lesson is not specific to artificial intelligence systems and what they can do. It is the old computer science lesson: do not build autonomously on interfaces that nobody promised to maintain.
Agentic systems work best when running against your own infrastructure and published APIs that have versioning contracts and backward compatibility promises that matter.