Hermes multi-agent team collaboration using fixed scripts for cross-agent communication

byǝɔɐǝԀʎzɐɹƆ •四月 24, 2026

0

Preface

From Openclaw to Hermes, I've attempted multi-agent team collaboration multiple times.

https://zelikk.blogspot.com/search?q=%E5%A4%9Aagent

Many times cross-agent communication fails, and the erroneous behavior is unpredictable — meaning the same prompt / agent.md file might work this time but fail on the next try.

The most recent case was using Hermes, and I had already previously prompted it to modify itself to gain the ability to bind agents to Telegram group topics, similar to Openclaw.

During usage,

Once, agent-design did not send a message to agent-watch but replied in its own topic instead

Two other times, when agent-watch sent messages to agent-design, they did not follow the specified message format.

https://www.youtube.com/watch?v=TBmbAcV6Gis

Approach

At the textual description level of SOUL.md, a sufficiently intelligent "person" or "model" should understand the effect I want.

* The SOUL.md before this experiment has been overwritten; you can refer to my configuration file on Openclaw.

https://github.com/crazypeace/openclaw-multi-agent-cooperate

However, it seems the current model (mimo-v2-pro 2026-4-20) does not yet meet my ideal of "intelligence" in team collaboration.

I've tried many different approaches and finally settled on a simple, straightforward method:

Agents only need to know they are working in a team and which other agents they collaborate with.
Agents do not know the team works within a Telegram group.
Agents do not know which topics other team members are bound to.
When an agent needs to send a message to another agent, it calls a designated script.
Inside the script, the Telegram bot token, group id, and topic id to use are hard-coded.

Suppose this script is /root/.hermes/profiles/agent-watch/scripts/send_to_agent-design.sh

Then agent-watch uses it like this:

scripts/send_to_agent-design.sh "message to send to agent-design"

Assume the team's group id is -1003311933525

agent-design's bound topic id is 168

Then, the pseudo-code for this script would be:

Get own bot token from own .env configuration file

Build a team message in the specified format
FROM agent-watch
TO agent-design
CONTENT "message to send to agent-design"

Use Telegram bot API's sendMessage method with parameters:
bot token = ( own bot token )
group id = ( -1003311933525 )
topic id = ( 168 )
message content =
( FROM agent-watch
TO agent-design
CONTENT "message to send to agent-design" )

Result

This works much better

Same task as before

Develop a random password generator page
Core functionality reference: https://crazypeace.github.io/xkcd-password-generator/
Visual style reference: https://onojyun.com/

The entire process had every message in correct format (of course, it's fixed in the script), and team messages were correctly sent to their respective topics.

It seems the model (mimo-v2-pro 2026-4-20) is capable of "understanding" the intent of sending a message to another team member, but at the execution level, this approach helps the model execute better without trying to invent novel methods.