O conteúdo está em inglês; a navegação e os filtros mantêm o idioma selecionado.

How to Use Agent “/goal” Workflows for Long-Running Tasks

Goal-oriented agent modes let an AI continue working across multiple turns instead of stopping after a single response. Effective use depends on a clear objective, measurable validation, and a defined stopping policy.

Publicado em 09/08/2026, 01:49:22Última atualização 09/08/2026, 01:49:22Publicação editorial

This workflow is suited to performance tuning, test repair, and technical experiments that require repeated measurement. A practical sequence is to establish a baseline, inspect the bottleneck, apply a change, and rerun the same test to determine whether the result improved. Using a separate worktree can make competing approaches easier to compare without changing the main branch.

The reported case involved optimizing large-video transcription in a CLI, with a specified model and a multilingual test video used for benchmarking. The author claims an improvement of more than two times. Available evidence supports the existence and general behavior of persistent goal modes, but it does not independently verify that performance figure, so the result should be treated as an anecdotal case study rather than a broadly established benchmark.

Delegating concrete implementation tasks to subagents may reduce reliance on an expensive coordinator model. Long-running sessions still need limits on budget, permissions, and runtime, along with human review; an agent’s own assessment that no further optimization is possible is not a reliable stopping rule by itself.

Evidências das fontes 6

A verificação de fontes não garante exatidão permanente. Considere a fonte original, a data e as correções posteriores.

持久目标 | Hermes Agenthermes-agent.nousresearch.com · Complementar

`↻ Continuing toward goal` `✓ Goal achieved` ## 致谢 `/goal` 是 Hermes 对 Ralph loop 模式的实现。面向用户的设计——跨轮次保持目标存活、不达成不停止，以及创建/暂停/恢复/清除控制——由 OpenAI Codex 团队的 Eric Traut 在 Codex CLI 0.128.0 中推广并落地。我们的实现是独立的（中央 `CommandDef` 注册表、`SessionDB.state_meta` 持久化、辅助客户端裁判、gateway 侧的适配器 FIFO 续行），但这个想法源自他们。功劳归于应得之人。 `/goal` `CommandDef` `SessionDB.state_meta` `/subgoal` [...] Hermes Agent Hermes Agent # 持久目标（`/goal`） `/goal` `/goal` 为 Hermes 设置一个跨轮次持续存在的目标。每轮结束后，一个轻量级裁判模型会检查目标是否已被助手的最新回复满足。若未满足，Hermes 会自动将一条续行 prompt（提示词）注入同一会话并继续工作——直到目标达成、你暂停或清除目标，或者轮次预算耗尽为止。 `/goal` 这是我们对 Ralph loop 的实现，直接受 Eric Traut（OpenAI）在 Codex CLI 0.128.0 的 `/goal` 中的启发。核心思路——跨轮次保持目标存活、不达成不停止——源自他们。此处的实现是独立的，并已适配 Hermes 的架构。 `/goal` ## 适用场景当你希望 Hermes 自主迭代、无需每轮重新提示时，使用 `/goal`： `/goal` `src/` `ruff check` 只需一轮即可完成的任务不需要 `/goal`。否则你需要说三次"继续" 的任务，才是它的用武之地。 `/goal` ## 快速开始 `/goal Fix every failing test in tests/hermes_cli/ and make sure scripts/run_tests.sh passes for that directory` 你将看到： [...] 💻 echo "4" > /tmp/