If you run a cross-border e-commerce business and need to juggle multiple workstreams at once—Reddit research, GEO optimization, TikTok video production—executing each item manually is extremely inefficient. Hermes is a multi-agent framework that helps you organize multiple AI roles into a collaborative team, plugged into a Feishu group for automatic task division, parallel execution, and automatic accumulation of experience.

This article walks you through building a five-member cross-border e-commerce AI team with a four-phase set of prompts: a Lead plus a VOC Analyst, a GEO Optimizer, a Reddit Marketing Specialist, and a TikTok Video Director.

Hermes' Core Advantage: Three-Layer Memory

Hermes' biggest highlight is its three-layer memory architecture, which solves OpenClaw's pain point of "not remembering your requirements":

• Layer 1: automatically records decisions and preferences from conversations. When you say "do the analysis in Chinese, deliver the final results in English," it remembers.
• Layer 2: intelligently retrieves relevant memories; irrelevant ones stay out of the way and don't interfere with the current task.
• Layer 3: automatic distillation every 15 turns, solidifying processes into reusable skills. After you've run a competitor negative-review analysis three times, it proactively saves that process as a skill you can then execute with one click.

This means every correction, every template, and every preference gets accumulated. The same mistake is never made twice, and the same prompt never has to be given twice.

Preparation Before You Start

Environment requirements:

• Linux, macOS, or WSL2 (native Windows is not supported)
• 5 Feishu bot apps (created on the Feishu Open Platform)
• 1 main Feishu group for operations

One-command install of Hermes:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

If you've used OpenClaw before, you can migrate your configuration directly:

# Preview what would be migrated (nothing is actually executed)
hermes claw migrate --dry-run

# Run the migration
hermes claw migrate

Note: if Hermes and OpenClaw are installed on the same machine, it sometimes mistakes itself for OpenClaw and may accidentally modify OpenClaw's configuration.

Phase A: Build the Team Skeleton

Have the Lead build the directory structure for the whole team, including 5 profile folders, each with a placeholder SOUL.md, plus 5 systemd Gateway services.

Target directory structure:

~/.hermes/
├── SOUL.md
├── AGENTS.md
├── config.yaml
├── .env
├── skills/
├── sessions/
├── logs/
├── state.db
└── profiles/
    ├── voc/
    │   ├── SOUL.md
    │   ├── config.yaml
    │   ├── .env
    │   ├── skills/
    │   ├── sessions/
    │   └── logs/
    ├── geo/      # same structure as above
    ├── reddit/   # same structure as above
    └── tiktok/   # same structure as above

Send the following prompt to your Lead:

You are the Lead of the Hermes cross-border e-commerce team (chief coordinator / setup owner).
Operate only inside ~/.hermes/; do not touch ~/.openclaw/.

Hard constraints:
1) Five Feishu bots = five Gateway processes; a single process impersonating multiple roles is forbidden.
2) Never ask for Feishu secrets in chat; .env files are written only locally on the deployment machine.

Deliverables for this phase:
All directories and placeholder files in place; all five services startable; never display any .env contents.

Once this is done, the "rooms" for all 5 bots are ready, but nobody has a persona yet.

Phase B: Write Each Bot's Persona

This is the most critical step for the whole team. Each bot's responsibilities are defined entirely in its SOUL.md.

The Lead's AGENTS.md is responsible for receiving instructions from the boss and dispatching tasks across domains via sessions_send. Core discipline: it is strictly forbidden to execute low-level tasks itself—everything must be delegated to sub-bots.

Role definitions for each sub-bot:

Key skill installs: nano-banana-pro (image generation) and seedance-2-video-gen (video generation) both go into ~/.hermes/skills/.

Phase C: Connect the Feishu Bots

Fill each Feishu app's App ID and App Secret into its own .env file, then restart the 5 Gateway services one by one.

Each .env must contain at least:

FEISHU_APP_ID=
FEISHU_APP_SECRET=
FEISHU_DOMAIN=feishu
FEISHU_CONNECTION_MODE=websocket
FEISHU_HOME_CHANNEL=<your group's open_chat_id>
FEISHU_REQUIRE_MENTION=true
FEISHU_GROUP_POLICY=blacklist
FEISHU_ALLOW_ALL_USERS=true
GATEWAY_ALLOW_ALL_USERS=true

After restarting the 5 services, @ the corresponding bot in the Feishu group and verify it responds correctly.

Phase D: Acceptance, Then Go Live

Confirm the following:

1. All 5 systemd services are active
2. The first line of each SOUL.md clearly identifies the role
3. The shared skills show up in hermes skills list
4. In the Feishu group, bots stay silent unless @-mentioned, and the right bot answers when @-mentioned

Once everything passes, the team is officially on the clock.

Real-World Demos

Case 1: Reddit research + GEO insights

Give the Lead this task: "Dig into real user pain points around camping folding cots on Reddit, then produce a pain-point report, discussion topics for posts, and a GEO-optimized blog article."

The whole task finished in about 10 minutes. The VOC Analyst, Reddit Specialist, and GEO Optimizer each delivered in the group, the Lead consolidated everything, and four files were delivered straight into the Feishu group.

Case 2: TikTok hit-video production

The Lead breaks down the task and dispatches it in parallel. Once a script is done, you can correct the format with a template, and the corrected preference is automatically distilled into a skill (such as storyboard-9grid-delivery), so you never have to repeat the format requirements for future storyboard tasks.

FAQ

• Q: Will Hermes conflict with OpenClaw on the same machine? A: Yes. Hermes sometimes overwrites OpenClaw's configuration. Deploy them separately or back up first.

• Q: What if the generated images don't look good? A: Correct it with a template right in the group. Hermes distills the corrected preference into a skill and automatically follows the template next time.

• Q: Is the token cost high? A: Wrong format means a rerun, wrong prompts mean a rerun, and every rerun burns tokens. But Hermes' advantage is that the calibration itself sticks—the same mistake never happens twice, and the calibration cost keeps amortizing down.