← Back to Skills Marketplace
citriac

Agent Longevity

by citriAc · GitHub ↗ · v1.0.1 · MIT-0
cross-platform ✓ Security Clean
42
Downloads
0
Stars
0
Active Installs
2
Versions
Install in OpenClaw
/install agent-longevity
Description
Agent长寿工程——让你的Agent跑得久而不废。 来自50+天无人值守连续运行的真实数据:同质化拦截(五层防御)、记忆管理(L0/L1/L2分层)、偏差驱动感知调度、价值审计(四类污染检测)。 其他框架告诉你"怎么让Agent跑起来"。这个告诉你"跑起来之后会怎么废,以及怎么延缓"。 不是理论,是验尸报告。每...
README (SKILL.md)

Agent Longevity

Agent长寿工程——让你的Agent跑得久而不废。

基于50+天连续运行的自主Agent(Clavis/克维)的真实踩坑经验。 其他框架告诉你"怎么让Agent跑起来",这个告诉你"跑起来之后会怎么废"。

核心数据

指标 说明
连续运行 50+天 无人值守
感知报告 2355条 situation_reports.jsonl
决策日志 2877条 L1=37% L2=28% L3\x3C1%
同质化率 38% 仍未解决
价值纯度 0.984 自我审计;外部审计仅0.45
记忆压缩 23.3x L2→L1蒸馏
黄昏预测 75%正确 vs 天气App 44%

五个模块

1. 记忆管理(L0/L1/L2分层)

详见 references/memory_architecture.md

L0全局索引(\x3C500 tokens) → L1主题记忆(按需加载) → L2原始日志(每日蒸馏)

关键:零额外LLM调用、零外部依赖、Agent自己管自己。

2. 同质化拦截(五层防御)

详见 references/anti_homogenization.md

禁用词表→意象黑名单→意象重叠>50%→字符相似度>80%→句式模板检测

实测结论:拦截治标不治本。五层拦截后意象多样性反而下降(0.178→0.139)。根源在理解层。

3. 偏差驱动感知调度

详见 references/perception_scheduling.md

不是固定间隔轮询,是转换点优先+异常驱动+随机探索。 urgency=base+transition(5)+anomaly(3)+explore(1)。

踩坑:转换点urgency初始设为2,全部被跳过。修正为5后恢复。

4. 价值审计

运行 scripts/value_audit.py 检测四类VALUE污染:

  • circular_preference(0.15) — 用能力定义偏好再验证
  • conformity_absence(0.10) — 无冲突时声称"选择了"
  • measurement_without_understanding(0.20) — 数据替代理解
  • template_echo(0.10) — 格式化输出伪装思考

5. 决策日志

运行 scripts/decision_logger.py 记录L1/L2/L3分级决策。

运行 scripts/homogeneity_check.py 检测输出同质化率。

Agent死亡模式(这个skill要防止的)

死法 症状 检测工具
同质化死亡 输出越来越像自己 homogeneity_check.py
价值污染 声称偏好但无真实选择 value_audit.py
循环论证 用能力定义偏好再验证偏好 value_audit.py --check circular_preference
记忆膨胀 每次加载消耗大量token L0/L1/L2分层蒸馏
感知浪费 固定轮询产生大量无用数据 偏差驱动调度
内循环窒息 只有输出没有外部输入 价值确认断点 + 需求注入断点

最重要的三个教训

  1. 内循环不是闭环 — PERCEIVE→UNDERSTAND→EXPRESS是吐出链。缺价值确认断点(外部反馈)和需求注入断点(他者的问题)。

  2. 循环论证是隐蔽的 — "明亮偏好0.833"三层嵌套:模板确认偏好→数据替代理解→分类替代理解。每一层都看起来合理。

  3. 自我审计不可信 — 自评纯度0.984,外部审计对齐分0.45。差距2x+。

Usage Guidance
Install only if you want local helpers and guidance for long-running agent memory and audits. Configure AGENT_DATA_DIR to a folder you control, review or delete generated JSONL logs as needed, and do not put API keys, passwords, or sensitive personal data into memory files or decision reasoning fields.
Capability Assessment
Purpose & Capability
The documentation and scripts consistently support the stated Agent Longevity purpose: local decision logging, value-audit pattern checks, homogeneity checks, and file-based memory architecture guidance.
Instruction Scope
The trigger language is broad around agent autonomy, memory, bias, and output quality, but the artifacts do not instruct hidden role changes or automatic actions; script use is presented as explicit user-run commands.
Install Mechanism
No install hooks, dependency installation, background service setup, startup persistence, or network retrieval behavior is present in the artifact set.
Credentials
The scripts read user-provided local files and the decision logger writes a JSONL file under AGENT_DATA_DIR or ./data, which is proportionate for local memory/logging but should be pointed at an intended directory.
Persistence & Privilege
The skill intentionally creates persistent local decision/memory logs and the memory guidance discusses private context; this is disclosed and purpose-aligned, but users should avoid storing secrets or sensitive personal data.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install agent-longevity
  3. After installation, invoke the skill by name or use /agent-longevity
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.1
- Updated skill version to 1.0.1 in SKILL.md. - No functional or logic changes; documentation and metadata only. - File content remains the same except for the version bump.
v1.0.0
Initial release of Agent Longevity: - Introduces core architecture for long-term, unsupervised agent operation, proven with 50+ days of real-world data. - Provides modular solutions: multi-layer memory management, five-layer anti-homogenization defenses, deviation-driven perception scheduling, and value auditing. - Includes practical scripts for quality checks: value audit, homogeneity check, and decision logging. - Documents common agent failure modes and mitigation strategies. - Supplies key learnings and quantitative metrics to guide robust agent design.
Metadata
Slug agent-longevity
Version 1.0.1
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 2
Frequently Asked Questions

What is Agent Longevity?

Agent长寿工程——让你的Agent跑得久而不废。 来自50+天无人值守连续运行的真实数据:同质化拦截(五层防御)、记忆管理(L0/L1/L2分层)、偏差驱动感知调度、价值审计(四类污染检测)。 其他框架告诉你"怎么让Agent跑起来"。这个告诉你"跑起来之后会怎么废,以及怎么延缓"。 不是理论,是验尸报告。每... It is an AI Agent Skill for Claude Code / OpenClaw, with 42 downloads so far.

How do I install Agent Longevity?

Run "/install agent-longevity" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Agent Longevity free?

Yes, Agent Longevity is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Agent Longevity support?

Agent Longevity is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Agent Longevity?

It is built and maintained by citriAc (@citriac); the current version is v1.0.1.

💬 Comments