← Knowledge Atlas · Concept
Feature Tracking(特性追踪)
特性追踪:结构化 JSON feature list,防止 one-shotting 和 premature victory
念
Feature Tracking
Feature Tracking — an external task list that blocks two failure modes
An externalized, machine-readable task list — a JSON feature list the agent can read for global progress and update completion against. It blocks both “one-shotting” (doing it all in one go) and “premature victory” (declaring done too early).
JSON structure
categorydescriptionstepspasses
authuser login flow3false
authtoken refresh2true ✓
…200+ entries…false
key constraints passes only the agent may only flip passes to true; it cannot delete or edit the description JSON, not Markdown models are less inclined to “casually” edit JSON structure, making it safer than Markdown end-to-end required a task is only marked passing after end-to-end test verification — self-judgment is forbidden
→ Long-Running Agents · Harness Engineering · Context ManagementAnthropic (2024)
Feature Tracking(特性追踪)
定义
Feature tracking 是 长时运行 agent 中用于追踪任务完成度的结构化机制。其核心是一份外部化的、machine-readable 的任务清单,agent 可以读取以了解全局进展,可以更新以记录自己的成果。
Anthropic 的实践
在 Anthropic 的 harness 设计 中,feature tracking 采用 JSON 格式的 feature list:
- 每条 feature 包含:
category、description、steps、passes(布尔值) - Initializer agent 基于用户需求生成完整列表(200+ 条),所有
passes初始为false - Coding agent 完成并验证一个 feature 后,将其
passes改为true
关键约束
- Agent 只能修改
passes字段,不得删除或编辑测试描述 - 选择 JSON 而非 Markdown,因为模型更不容易”顺手”修改 JSON 结构
- Feature 必须经过端到端测试才能标记为通过
设计意图
Feature tracking 同时解决两个 长时 agent 的失败模式:
- 防止 one-shotting:明确的 feature list 让 agent 知道有多少工作要做,逐个推进
- 防止 premature victory:未通过的 feature 是客观证据,agent 无法”觉得差不多了”就停下
相关概念
- Long-running agents — feature tracking 服务的场景
- Harness engineering — feature tracking 是 harness 的组成部分
- Context management — feature list 是跨 session 状态传递的关键载体
References
sources/anthropic_official/effective-harnesses-long-running-agents.md