AI Engineering · LLM applications that ship
How do agents fail on multi-step tasks?
Short answer
Agents compound small errors: wrong tool choice, bad arguments, ignored observations, loops and premature stopping. Each step's error rate multiplies across the trajectory.
A five-step task where each step succeeds eighty percent of the time succeeds only about a third of the time end to end. Agents also drift from the goal, repeat a failing action, or declare success without checking the result. The longer the trajectory, the more the failure modes matter.
- Wrong tool or wrong arguments, often from overlapping schemas.
- Observation ignored: the agent repeats an action that already failed.
- Looping: no progress but continued spending.
- Premature completion: the agent stops before verifying the outcome.
- Context loss: earlier constraints fall out of the window.
Follow-ups you should expect
Interviewers use these to check whether the first answer was understanding or memorisation.
- 1How would you detect an agent loop automatically?
- 2What would you add to force verification before completion?
- 3How would you evaluate an agent's trajectory, not just its final answer?
Say it in the interview
Agents fail by compounding small mistakes. A wrong tool call changes the state, the next step reasons from bad information, and by step five the task is off track. I see loops, ignored errors, premature completion and context loss most often. I log every step with inputs and outputs, cap iterations, and require a verification step before the agent can declare success. Evaluation has to look at the trajectory, not only the final message.
Use this as a model, not a script. Saying it in your own words is what makes it survive follow-up pressure.
Related questions
Reading is step one
Now say it out loud.
Understanding an answer and delivering it under pressure are different skills. Rehearse this topic in a live AI Engineering mock interview and get scored feedback on depth, clarity and confidence.