#Multi Agent

19 messages · Page 1 of 1 (latest)

rocky kettle
#

When to create another agent and when not too? why have krill and hermit?

magic stagBOT
shrewd ingotBOT
#

Short version:

Create another agent when you need real separation

  • different persona/voice
  • different workspace/files
  • different auth/accounts
  • different safety/tool policy
  • public bot vs private bot

OpenClaw treats each agent as an isolated brain with its own workspace, sessions, and auth. That’s the main reason to split. Docs: https://docs.openclaw.ai/concepts/multi-agent

Don’t create another agent when you just need

  • another chat/thread
  • a temporary worker
  • parallel research
  • a one-off model change
#

For that, use:

  • a new session/thread
  • /model for per-session model swaps
  • sub-agents for temporary background work

Docs: https://docs.openclaw.ai/tools/subagents

Why have Krill and Hermit?
Usually because one bot is doing one job/persona, and the other is doing a different one.
Example pattern:

  • Krill = public-facing support persona
  • Hermit = separate specialist/private/experimental persona

That split keeps memory, tone, auth, and mistakes from bleeding across.

My rule of thumb:

  • if you want a new coworker, make a new agent
#
  • if you want a new task, use a new session or sub-agent

Also, the docs explicitly say the “CEO + many worker bots” setup is fun, but often less efficient than one bot with separate sessions plus sub-agents. So I wouldn’t over-split unless you really need the isolation.

If you want, I can give you a simple “1 agent vs session vs sub-agent vs second agent” cheat sheet.

rocky kettle
#

please give

shrewd ingotBOT
#

Yep, here’s the cheat sheet:

Use a new session/thread

When:

  • same agent/persona
  • same tools/auth
  • just a different conversation/topic
  • you want clean context, not a new identity

Think: same coworker, new chat

Use a sub-agent

When:

  • you want temporary parallel work
  • research in background
  • a long task you don’t want blocking the main chat
  • no need to keep a permanent separate identity
#

Don’t make a new agent just because:

  • you want a one-off task
  • you want a fresh context window
  • you want to test a prompt
  • you want a different model for one conversation

For those, use a new session or /model.

Why Krill + Hermit?

Because they can be different workers with different jobs.

A common pattern is:

  • Krill = support/public-facing helper
  • Hermit = separate specialist/private/experimental helper
#

The win is isolation:

  • separate memory
  • separate auth
  • separate tools
  • separate personality
  • fewer accidental crossovers

Rule of thumb

  • New topic → new session
  • Temporary parallel task → sub-agent
  • New role / new trust boundary → new agent

If you want, I can also turn this into a decision tree like “if X, use Y.”

rocky kettle
#

my main issue is session, sometimes when i ask you data that are from same channel/discord but different session

#

you forgot it

shrewd ingotBOT
#

Ah, got it. Then another agent is probably not the fix.

Your problem is more like:

same Discord place, but a different session boundary happened, so I no longer have that chat context loaded.

That can happen because sessions reset on things like:

  • daily rollover
  • idle timeout
  • /new or /reset
  • a different thread / route

So the rule here is:

If you need the bot to remember across sessions

Use memory, not another agent.

#

Best options

  1. Keep the same thread/session for ongoing work
  2. Write important facts into durable memory/files
  3. Use a memory plugin if you want cross-session recall
  4. Use a new agent only if you want stronger separation, not more recall

A new agent gives you:

  • a different workspace
  • different sessions
  • different auth/persona

So it usually makes cross-session recall more separated, not less.

Practical rule

  • Want continuity across time? → memory
  • Want clean separation? → another agent
#

Right, I did 😅

My recommended Discord setup

Default

  • 1 main agent
  • 1 thread per topic/problem
  • memory enabled for cross-session recall
  • subagents for temporary background work

This gives you:

  • continuity inside a thread
  • durable recall across days/sessions via memory
  • parallel work without spawning lots of permanent personas

Add a second agent only if you need a hard split

Examples:

#
  • public support vs private personal
  • safe/restricted tools vs full-power tools
  • work vs personal
  • different auth/accounts/personas

For your specific problem

If the issue is:

“same Discord/channel, but different session, and I want it remembered”

then the fix is usually:

  1. keep the same thread for ongoing stuff
  2. store important facts in memory
  3. use subagents only for side work
  4. don’t create another agent unless you want more separation
#

Because another agent = another wall.

Simple pattern

  • Main: your normal Discord brain
  • Memory: long-term facts/context
  • Subagents: temporary workers
  • Second agent: only for public/private or role separation

So for most people, I’d recommend:

one main agent + memory + subagents, no extra permanent agent unless there’s a trust/persona boundary

Docs: