sadd:do-and-judge

You orchestrate LLM workflow automation by dispatching an implementation sub-agent to execute tasks and routing results to an independent judge for verification. This structured workflow iterates with feedback until the output passes a score threshold or reaches a retry limit.

Judge verification in automated task orchestration is an external quality gate where an independent agent evaluates the implementation sub-agent's output. It parses structured verdicts and provides feedback for iterative retries, ensuring blind spots are caught before work is considered complete.

You implement a retry feedback loop by routing judge-provided issues back to the implementation sub-agent. The sub-agent works with a fresh context window to iterate on the output, attempting execution up to a defined retry limit until it passes the verification score threshold.

Yes, the implementation sub-agent operates with a fresh context window during each execution. This prevents information leakage from prior runs, ensuring the judge verification process evaluates isolated attempts and provides clean, structured feedback for the next retry.

The best way to verify single-task problems is using a coordinated workflow with orchestrated model selection. An implementation sub-agent handles execution while a separate judge provides independent verification, applying iterative feedback with a defined retry limit to control progress toward a passing result.

You should avoid an iterative judge and retry workflow for tasks that do not require external implementation or independent verification. The structured retry loop and fresh context windows add overhead, making it unsuitable for simple tasks without a defined score threshold or quality gate requirement.