5K

Reflection Loop

Scores each draft against a fixed rubric, feeds the critique into a revision, and repeats until the draft passes a threshold or the retry budget ends

established Updated 2026-10-13 By Nikola Balic (@nibzard) Based on Shinn et al. (2023)
View .md Edit on GitHub
Add to Pack
or

Saved locally in this browser for now.

Cite This Pattern
APA 
Nikola Balic (@nibzard) (2026). Reflection Loop. In *Awesome Agentic Patterns*. Retrieved October 14, 2026, from https://agentic-patterns.com/patterns/reflection
BibTeX 
@misc{agentic_patterns_reflection,
  title = {Reflection Loop},
  author = {Nikola Balic (@nibzard)},
  year = {2026},
  howpublished = {\url{https://agentic-patterns.com/patterns/reflection}},
  note = {Awesome Agentic Patterns}
}

Use when

  • Output must meet explicit quality criteria
  • Single-pass answers miss edge cases or constraints
  • Writing, reasoning, or code tasks with a clear scoring metric

Avoid when

  • No well-defined metric to score drafts
  • Extra compute per answer is not acceptable

Problem

Single-pass generation frequently misses edge cases, constraints, or quality criteria that become obvious on review. Without a structured revision loop, agents return the first plausible answer even when a better answer is reachable with lightweight critique.

Solution

After generating a draft, run an explicit self-evaluation pass against defined criteria and feed the critique into a revision attempt. Repeat until the output clears a threshold or retry budget is exhausted.

Use stable scoring rubrics (correctness, completeness, safety, style) so the loop improves objective quality rather than free-form restyling. For reduced bias, use a separate model for critique (dual-model architecture) at the cost of additional compute.

for attempt in range(max_iters):
    draft = generate(prompt)
    score, critique = evaluate(draft, metric)
    if score >= threshold:
        return draft
    prompt = incorporate(critique, prompt)

How to use it

Use this when quality must meet explicit criteria in writing, reasoning, or code generation. Keep loop budgets small; there is no universally established optimum, so log score deltas per iteration and stop when the deltas flatten rather than assuming a fixed iteration count. Note that 2025-2026 reasoning models already internalize a form of self-checking in their thinking, so hand-rolled reflection scaffolds pay off mostly with non-reasoning models or when the rubric is domain-specific.

Trade-offs

  • Pros: Improves outputs with little supervision.
  • Cons: Extra compute; may stall if the metric is poorly defined.

References