Five ways a beginner's agent loop breaks
Most beginner agents I've come across look something like this: while True: reply = llm(messages, tools) if reply.tool_call: result = run_tool(reply.tool_call) messages.append(result) else:
Most beginner agents I've come across look something like this:
while True:
reply = llm(messages, tools)
if reply.tool_call:
result = run_tool(reply.tool_call)
messages.append(result)
else:
return reply.text
It runs fine in a demo, then fails in five very predictable ways.
1. No clear stop condition
The model keeps making tool calls because nothing explicitly tells it when to stop. Fix: set a step limit and add a dedicated "finish" tool, so stopping becomes an intentional decision by the model.
2. Repeating the same failed call
A tool returns an error, and the model tries again with exactly the same arguments. Fix: feed the error back unchanged and track repeated calls. After three identical attempts, stop and replan, or ask the user.
3. Errors that disappear
run_tool catches an exception and returns "". The model continues as if nothing happened and often makes up a result. Fix: treat errors as data. Return them as readable text so the model can understand and respond to them.
4. Context that keeps expanding
Every tool result gets added in full, and by step 20 the original instructions are lost in the noise. Fix: shorten or summarize earlier results, while keeping the actual task visible at the top.
5. Claiming "done" without checking
The model announces that the task is finished, but nobody verifies it. Fix: add a verification step before accepting the answer: run a test, execute a query, or have another prompt review the result.
I've been working on Way of the Agent, a game-based course about agent engineering, and these five problems appear as daily exercises: a quick scenario - "your agent keeps repeating the same tool call", "a tool call fails, how should your loop respond?" - followed by four possible solutions to pick from. Choosing the correct solution turned out to test understanding better than answering questions about definitions.
You can start for free, with a subscription unlocking the complete path, available on web, Android and iOS, in 8 languages: https://wayofagent.app
What failure would you put at number six? I'm gathering real-world examples.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.