If your plugin starts repeating questions, forgetting what the user said, or hallucinating field values after a few turns, the config probably isn't the problem. The context window is.
Every plugin has a context budget: the total information the AI can hold in working memory during a conversation. Slot descriptions, action outputs, system instructions, conversation history. They're all competing for the same finite space. And here's the thing: before your user says a single word, your plugin has already burned thousands of tokens just loading up.
Every turn after that is a fight for what's left.
This is why plugins that work fine in testing fall apart in real conversations. Short demos never push the budget. Real users do.
The good news: most context waste comes from just two places. Overstuffed slot descriptions and action outputs returning way more data than the AI actually needs. Clean those up and your plugin can hold a 20-turn conversation without losing the thread.
Want to go deeper? The full breakdown is in the latest issue of the Agent Studio Developer Newsletter. Find it in Community or subscribe to get it in your inbox.
📬 Subscribe to the Agent Studio Developer Newsletter in Community or via email.