Check the input the agent actually received — not the input you think you sent. Most agent misbehaviour starts with the agent receiving different, incomplete, or malformed data from what you intended.
Why the Input Is Always the First Stop
When your agent does something unexpected, the natural instinct is to jump to the middle of the trace and look for the moment it went wrong. That is like arriving at a car crash and immediately examining the airbags before checking what the driver was doing. Start at the beginning.
The vast majority of agent misbehaviour traces back to one of three input problems: the agent received the wrong data (a variable that did not populate correctly), incomplete data (a field that was supposed to be there but was empty), or data in the wrong format (a date string where a number was expected, or a line break in the middle of a value that broke the JSON). Fix the input, and many apparent logic failures disappear entirely.
How to Check the Input in a Trace
In Claude, the very first message in the trace array is the system prompt plus the initial user message. Read those two carefully. Does the system prompt look the way you designed it? Are all the variables filled in correctly — student name, lesson title, context window? Is the user message exactly what the student sent, or has something been truncated or escaped?
In n8n, click the trigger node first. Expand the output panel and look at the data structure. Is every field present? Are the values correct types? In Zapier, open the first step in the task history and check the data the trigger handed to the next step. A missing field at step one will cause every subsequent step to behave incorrectly, and the error will only show up several steps later — making it look like the AI or the final action was the problem when it was actually the input data that was broken.
Once you confirm the input was correct, move to the first tool call. Check what parameters the agent passed to the tool and whether those match the tool’s expected format. If both the input and the first tool call look right, keep moving forward through the trace one step at a time.
What This Means for Educators
Campus AI agents often pull data from multiple sources: a student’s FluentCRM record, a lesson from FluentCommunity, a question from a form submission. Any one of those data sources can produce bad input — a field left blank in a CRM record, a lesson with a special character in the title, a form that sends an empty string instead of null. Your agent’s instructions might be perfect, but if the data coming in is broken, the output will be wrong.
Training yourself to check the input first saves significant debugging time. In most cases, you will find the problem in the first thirty seconds without having to read the entire trace.
The Simple Rule
Garbage in, garbage out — and garbage usually shows up at the very start of the trace. Read the input before you read anything else. If the input was clean, work forward one step at a time. Resist the urge to start in the middle.