sadd:do-and-judge

Multi-agent orchestration automates task quality verification by dispatching a meta-judge to generate evaluation criteria and an implementation agent to execute the task. A judge agent then evaluates the output against the criteria, enforcing iterative retries until the quality gate is passed.

Multi-agent orchestration with an LLM judge coordinates a meta-judge, an implementation agent, and a judge agent to complete and verify tasks. It applies to tasks requiring iterative refinement and safe gating before acceptance, using fresh-context execution per attempt.

Iterative refinement for automated task execution is set up by dispatching a meta-judge and an implementation agent in parallel. A judge agent then evaluates the output, and the system repeats execution with feedback up to a fixed retry limit until criteria are met.

Multi-agent task execution works without external dependencies because it operates as a self-contained quality gate. It uses fresh-context execution per attempt to avoid context bleed and dispatches sub-agents internally to complete and verify the work.

Automated judging can be used for small coding tasks by providing a task description to the orchestration system. It returns a verified implementation plan by dispatching sub-agents and enforcing a quality gate before acceptance.

The retry limit when using an LLM judge for task gating is a fixed number of retries enforced by the judge agent. If the implementation fails to meet the meta-judge's criteria, the system repeats with feedback until it passes or the limit is reached.