Loop Engineering, or Why Your Agent Goes in Circles

Strip away the marketing and an agent is this:

while not done:
    action = model(context)
    result = execute(action)
    context = update(context, action, result)

Loop engineering is the study of that while. It sounds trivial. It's where most agents fail.

The classic failure modes

1. The infinite retry. Tool fails, model tries the exact same call, tool fails, repeat. Fix: detect duplicate calls and inject "that failed twice; try a different approach or stop."
2. The runaway plan. Model decides it needs 40 sub-steps for a 2-step task. Fix: budgets (turns, tokens, dollars) enforced by the harness, plus a "checkpoint: is this still on track?" prompt every N turns.
3. Premature victory. Model declares success without verifying. Fix: require evidence. "Done" must be accompanied by a test run, a diff, a URL that returns 200.
4. Context bloat. Every tool result gets appended until the window is full of stale logs. Fix: compaction. Replace old tool results with one-line summaries once they're consumed.
5. Lost goal. After 30 turns the original task has scrolled out of view. Fix: pin the goal. Re-inject the task statement near the end of the window each turn.

Loop shapes that work

• ReAct (reason → act → observe): the default. Fine for short tasks.
• Plan-then-execute: generate a plan up front, execute steps, re-plan only on failure. More predictable, less adaptive.
• Reflexion: after a failure, ask the model to write a lesson, then retry with the lesson in context. Surprisingly effective for coding.
• Critic loops: a second model (or the same model with a different prompt) reviews the output before it's accepted. Cheap insurance.
• Human-in-the-loop checkpoints: the loop pauses at defined points (before spending money, before deleting things) and waits.

State: what the loop remembers

The loop needs an explicit state object, not just the message history. At minimum:

• the original goal, verbatim
• the current plan and which step you're on
• facts established so far (file paths, IDs, decisions)
• what has been tried and failed

Keep it small and structured. Regenerate the model's context from this state each turn rather than letting history accumulate.

A test you should run

Give your agent a task that is impossible (a file that doesn't exist, an API that always 500s). A good loop notices within a few turns, reports clearly, and stops. A bad loop burns your budget trying forever. This one test finds more bugs than any benchmark.