跪拜 Guibai
← All articles
Agent · OpenAI · AI Programming

A Five-Agent AI Team Runs Cross-Border E-Commerce Ops from Feishu

By 饼干哥哥 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Hermes' three-layer memory turns agent corrections into persistent skills, which addresses the core reliability problem that made multi-agent workflows fragile in earlier frameworks like OpenClaw. For anyone running agent teams that need to survive long task chains and session boundaries, this architecture means the break-in cost is a one-time investment rather than a recurring tax.

Summary

A full cross-border e-commerce AI team—Lead, VOC analyst, GEO optimizer, Reddit marketer, and TikTok director—was rebuilt on the Hermes agent framework after an earlier OpenClaw version went viral. The team runs autonomously: the Lead decomposes tasks, dispatches them to specialist agents, and consolidates deliverables directly in a Feishu group. In one case, the team scraped Reddit for user pain points and produced a GEO-optimized blog post in 10 minutes. In another, it generated a nine-grid TikTok storyboard, accepted a format correction mid-task, and persisted that correction as a reusable skill.

The build process follows a four-stage prompt sequence that constructs the directory skeleton, writes each agent's persona into SOUL.md files, connects five separate Feishu bots via Gateway processes, and runs an acceptance check. Hermes' three-layer memory architecture—automatic preference recording, intelligent retrieval, and periodic skill solidification—means corrections compound into a growing toolbox. The real cost is the break-in period: every rerun burns tokens, but Hermes retains those lessons so the same mistake never happens twice.

Takeaways
Five specialized AI agents—Lead, VOC analyst, GEO optimizer, Reddit marketer, and TikTok director—coordinate entirely through a Feishu group chat, with each agent running as a separate Gateway process.
A Reddit-to-GEO research task completed in 10 minutes, producing a pain-point report, post topics, and an optimized blog article.
A TikTok storyboard task accepted a mid-stream format correction, and the agent persisted the corrected nine-grid template as a reusable skill named storyboard-9grid-delivery.
Hermes uses a three-layer memory system: automatic recording of preferences, intelligent retrieval of relevant memories, and automatic skill solidification every 15 conversation rounds.
The four-stage build process uses structured prompts to create directory skeletons, write agent personas into SOUL.md files, connect five Feishu bots, and run acceptance checks.
Each agent's SOUL.md defines hard boundaries—the VOC analyst cannot write GEO copy, and the Reddit expert cannot produce TikTok videos—preventing role bleed.
Corrections and preferences compound into skills over time, so the same format or prompt mistake never needs to be corrected twice.
The framework supports migrating existing OpenClaw configurations with `hermes claw migrate`, including a dry-run preview mode.
Conclusions

The jump from OpenClaw to Hermes in just three months shows how fast agent memory architectures are evolving—session persistence and skill solidification are becoming table stakes, not differentiators.

Persisting a mid-task format correction as a reusable skill is a concrete example of agent learning that compounds; the system doesn't just remember, it operationalizes the memory so future tasks skip the correction step entirely.

The four-stage prompt sequence effectively treats the Lead agent as a DevOps engineer that provisions infrastructure, writes configs, and runs acceptance tests—blurring the line between agent orchestration and infrastructure-as-code.

The explicit role boundaries in each SOUL.md file prevent the common multi-agent failure mode where agents overstep their remit, but this rigidity also means the system's quality depends entirely on how well those boundaries are initially defined.

Concepts & terms
Hermes three-layer memory
An agent memory architecture with three tiers: automatic recording of conversational decisions and preferences, intelligent retrieval of only relevant memories to avoid context pollution, and periodic automatic refinement (every 15 rounds) that solidifies repeated workflows into reusable skills.
GEO (Generative Engine Optimization)
Optimization for AI-powered search engines like Perplexity and Google SGE, which differs from traditional SEO by requiring quantitative data, authoritative citations, and natural language structure rather than keyword density.
SOUL.md
A configuration file in Hermes that defines an agent's core identity, responsibilities, hard boundaries (what it must never do), and deliverable formats—functioning as the agent's immutable persona and job description.
Skill solidification
Hermes' process of automatically converting repeated workflows or user corrections into persistent, callable skills, so that a format fix or preference stated once becomes a permanent capability rather than requiring re-instruction.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗