Generative Loops

A generative loop is any orchestration shape where an agent’s output feeds back into another round of generation. This page is about the kind that includes review: a generator/reviewer loop, where one agent produces work and a reviewer agent evaluates it and sends it back for revision. Pairing a generator with a reviewer only improves quality if the loop has a defined stopping point and a real way to measure whether an iteration improved. Without both, it either runs forever or settles on something nobody checked.

Key takeaways

• Cap the loop and define “converged.” Without both, a loop runs forever or settles on something nobody checked.
• Use a review loop only when needed. It’s the one pattern that can run indefinitely, so it’s the costliest to get wrong.
• Send capped runs to human review. Hitting the iteration limit isn’t the same as converging.

The problem

Agentic Workflows names four orchestration patterns from the design-system-ops orchestration guide (knowledge-notes/agent-orchestration-guide.md): sequential chain, parallel agents, supervisor, and what the guide calls the feedback loop. This wiki uses the more specific name, generator/reviewer loop.

The first three each have a built-in end. A sequential chain finishes when its last agent runs. Parallel agents finish when every branch reports back. A supervisor finishes when it stops delegating. None needs a stopping rule, because stopping is just finishing. A generator/reviewer loop is the one pattern that routes work back to itself, and nothing about that shape says when to quit. Per the same guide, it “can loop indefinitely if convergence criteria are not defined.”

Practices

Cap the loop and define “converged”

A generator/reviewer loop needs an explicit stopping condition before it runs, not a hope that it’ll settle. The design-system-ops orchestration guide writes it as a rule: “The loop terminates when the reviewer reports no findings above a configured severity threshold, or when the maximum iteration count is reached.” It recommends three iterations at most. Three isn’t magic. An unbounded loop is a sign of a convergence rule nobody wrote down. Their broader rule applies too: “never silently skip a failed step.” A loop that never converges is a hidden failure, not a thorough one.

Use a review loop only when the task needs one

Romina Kavcic’s test for whether a task deserves an agent at all, “find the simplest solution possible, and only increase complexity when needed,” applies to which pattern you choose, too. A generator/reviewer loop is the most expensive of the four to get wrong: it’s the only one that can run indefinitely, and every extra round costs more agent time. If the work is predictable and doesn’t need review, a sequential chain or parallel agents do the job without the open-ended shape. The orchestration guide keeps the loop for “quality-critical output where a single pass is insufficient,” and names component descriptions, documentation, and governance rules as the work that benefits.

Measure whether a format works

Diana Wolosin’s benchmark of eight MCP configurations against 1,056 prompts (propose a format, measure how agents resolve it, revise) is a generative loop applied to infrastructure: a format-evaluation loop, where the candidate is a context format, not a piece of content. It isn’t one of the four orchestration patterns. It’s evidence that the same discipline (define “correct,” then measure against it) recurs outside workflow design. A format that looks reasonable isn’t necessarily the one that performs best. Context engineering has the full account.

Ship the loop as a UI pattern

AWS Cloudscape treats response feedback as a shippable component, not an internal evaluation step: thumbs-up and thumbs-down controls, plus a dedicated response regeneration pattern for when an output misses. It’s the same discipline as the agent-level loop, applied where a real user meets generated output.

Common mistakes

• Shipping whatever the last iteration produced. Hitting the cap isn’t the same as converging. The guide’s rule: “If the loop reaches the maximum without converging, the output is flagged for human review.”
• Reaching for a generator/reviewer loop when a simpler pattern would do. Having no built-in stopping point is a cost, not a feature. It’s only worth paying when the task truly needs review and revision rather than a fixed sequence of steps.