Building the AI agent is easier than keeping its live data reliable
I have been realizing lately that the hardest part of AI workflows is not even the model any more.
Getting an agent to work is one thing. Keeping the live data behind it stable over time is a totally different problem.
It is okay at first and then bit by bit the little things add up. Rate limits, broken scraping jobs, outdated context, retries, duplicate data, random platform changes, monitoring issues it never really ends.
I have found this to be particularly messy when the workflow is live conversations from places like Reddit or X. One bad data source can mess up the whole workflow pretty fast.
The AI models seem to be improving much more quickly than the infrastructure around live data and integrations.
Curious if other people building AI agents are running into the same thing or if people have found better ways to manage it now?