The agent loop is the iterative cycle at the heart of every agent: decide, act, observe, repeat — continuing until the goal is met, a stopping condition is hit, or a safety limit is reached. The observation from each step is what feeds back into the next decision, which is why this single loop covers what's sometimes described separately as a "feedback loop."
The Loop, Explicitly
iteration = 0
max_iterations = 10 # safety limit — never omit this in production
while iteration < max_iterations:
decision = llm.decide_next_step(goal, context, history)
if decision.type == "final_answer":
return decision.content
result = execute_tool_safely(decision.tool, decision.arguments)
history.append({"action": decision, "observation": result})
iteration += 1
# Loop exited without a final answer — this needs explicit handling,
# not a silent failure
return escalate_or_report_incomplete(history)Why the Iteration Cap Is Not Optional
Without a hard maximum, a model that gets stuck — repeating a failing tool call, oscillating between two unhelpful actions — can loop indefinitely, burning cost and time with no progress. See Infinite Agent Loops. A production agent loop always has an explicit exit condition beyond just "the model said it's done."
What "Feedback" Means Here
Each observation becomes part of the context for the next decision — this is what lets an agent adapt: if a tool call fails or returns unexpected data, that information is available to the LLM when deciding its next move, rather than the agent blindly continuing a fixed plan regardless of what actually happened.
Practical Use Case
A data-analysis agent's second query is only decided after seeing the first query's results — the loop is what makes that adaptive, multi-step behavior possible, as opposed to a fixed workflow that runs the same two queries regardless of what the first one returns.
Common Mistakes
- No maximum iteration count, or one so high it doesn't meaningfully protect against runaway cost
- Not persisting the loop's history/state incrementally, so a crash mid-loop loses all progress instead of allowing resumption from a checkpoint (see State Checkpoints)
- Silently returning nothing useful when the iteration cap is hit, instead of surfacing that the agent failed to complete the task
Interview Relevance
"How do you prevent an agent from looping forever?" is a very common practical question — a strong answer covers hard iteration limits, detecting repeated/unproductive actions, and timeouts, not just "hope the model stops on its own."
Practice Question
Write pseudocode for detecting that an agent has called the exact same tool with the exact same arguments three times in a row, and should stop instead of continuing.
