The iterative refinement pattern uses a looping mechanism to progressively improve an output over multiple cycles.1 Like review and critique, it is an implementation of the loop pattern.
Diagram
Open the interactive version here to pan and zoom.
How it works
One or more agents work inside a loop, modifying a result that’s stored in the session state on each iteration. That shared state is the distinguishing mechanic — the artifact persists across cycles and each pass edits it rather than regenerating from scratch.
The process continues until the output meets a predefined quality threshold or reaches a maximum number of iterations. The iteration cap isn’t optional bookkeeping; it’s what prevents the infinite loop that the base loop pattern warns about.
When to use it
Use it for complex generation tasks where the output is difficult to achieve in a single step — writing and debugging a piece of code, developing a detailed multi-part plan, drafting and revising a long-form document.
The source’s example is a creative writing workflow: an agent generates a draft of a blog post, critiques the draft for flow and tone, then rewrites it based on that critique. The loop repeats until the work meets a predefined quality standard or hits the maximum number of iterations.
From the pattern-comparison guidance, the workload profile is:
- open-ended or complex generation tasks that are difficult to complete in a single attempt,
- the agent must progressively improve the output over multiple cycles,
- no need for model orchestration,
- output quality is prioritized over latency.
Note that this is the one pattern the source lists in two categories — under deterministic workflows and under workflows involving iteration. Both are correct: the loop itself is code-driven and predictable, while the generation happening inside it is open-ended. See Choosing an Agentic Design Pattern.
Trade-offs
The pattern produces highly complex or polished outputs that would be difficult to reach in a single step.
The looping mechanism increases latency and operational cost with every cycle, and it adds architectural complexity: exit conditions — a quality evaluation, a maximum iteration limit — have to be designed carefully to prevent excessive cost or uncontrolled execution.
Footnotes
-
Samantha He, “Choose a design pattern for your agentic AI system”, Google Cloud Architecture Center, last reviewed 2026-05-28. Local copy:
references/google-cloud-agentic-design-patterns.md. ↩