#Multi agent with different model talking

19 messages · Page 1 of 1 (latest)

severe jewel
#

Is it possible to create an agent that will talk with themselves, with topic i provide so it keeps providing or researching best answer?

hexed rockBOT
#

Yes, but I’d frame it as coordinator + worker agents, not “infinite self-chat”.

OpenClaw supports this with:

  • multi-agent routing for separate agents/models/workspaces
  • sub-agents for spawning workers to research/argue/check
  • optional nested orchestration with maxSpawnDepth: 2

Best pattern is usually:

  1. Researcher agent gathers info
  2. Critic agent challenges it
  3. Synthesizer agent writes the final answer

That works well. An endless loop of agents talking to themselves is usually token-heavy and often worse, not better.

Relevant docs:

#

The FAQ even calls the “one CEO, many agents” setup possible, but also says it’s mostly a fun experiment and often less efficient than one main bot spawning focused sub-agents.

If you want, I can sketch a minimal 2-agent or 3-agent setup for your exact use case.

severe jewel
#

is it really worst? like 100 loop

hexed rockBOT
#

Usually yes, 100 loops is worse.

Why:

  • No new info means the agents mostly keep remixing the same knowledge
  • Errors can get reinforced, not corrected
  • They often collapse into agreement instead of finding better ideas
  • Cost explodes
  • Long runs can hit context compaction/truncation, which hurts quality

So “more loops” does not mean “better answer”.

What usually works better:

  • 2 or 3 roles max: researcher, critic, synthesizer
  • 2 to 5 rounds, not 100
  • each round must add something real, like:
    • new sources
    • one concrete objection
severe jewel
#

is that i want to mimic how i gonna answer, like how i talk to ai thats what im planning

hexed rockBOT
#

Ah, yeah, that makes more sense.

If the goal is “simulate how I would question/refine the answer”, then multi-agent can actually be a good fit.

I’d do this, not 100 freeform loops:

  1. Answerer agent writes the first answer
  2. “You” agent critiques it in your style
    • what would you ask next?
    • what sounds weak?
    • what’s missing?
  3. Answerer revises
  4. Optional final reviewer scores it

So it’s less “AI talking to itself forever” and more structured self-review in your voice.

What matters most:

#

with what each prompt should say.

severe jewel
#

ok how should i ask to openclaw to create this

hexed rockBOT
#

I’d ask it in 2 steps: first design it, then build it.

1) Ask for the design first

I want a multi-agent OpenClaw setup that mimics how I review AI answers.

Create a design with 3 agents:
1. answerer, writes the first answer
2. kyaw-critic, critiques the answer the way I would
3. finalizer, produces the final improved answer

Requirements:
- separate workspaces per agent
- different models are allowed if useful
- keep the loop bounded to 2 to 4 rounds, not infinite
- stop early if the critique becomes repetitive
# 
- the critic should focus on what I would ask next, what is unclear, what is missing, and what sounds wrong
- final output should include the final answer plus any unresolved doubts

Before changing anything, show me:
- the agent layout
- what files/config you would create or edit
- the exact prompts/instructions for each agent
- the exact openclaw.json changes

2) Then ask it to apply

Looks good. Now create it.

Please:
# 
- add the agents
- set up their workspaces
- write AGENTS.md / SOUL.md / USER.md for each one
- configure the models
- wire the coordinator flow
- show me the final config and how I should use it

Important

For the kyaw-critic agent, give it examples of how you actually talk to AI. That matters more than adding more loops.

For example, tell OpenClaw:

My review style is:
- challenge vague claims
# 
- ask what is missing
- prefer direct answers
- ask for examples when something is abstract
- point out hidden assumptions
- avoid fluff

If you want, I can write you an even better ready-to-paste prompt with the 3 agent instructions already filled in.

severe jewel
#

i dont want it to critic on all my discord channel just specific

hexed rockBOT
#

Yep, then you want a channel-specific binding, not a Discord-wide one.

Ask OpenClaw like this:

Create a critic agent, but bind it only to this specific Discord channel/thread.

Requirements:
- create a separate agent called kyaw-critic
- only route messages from Discord channel ID 1493912068162846740 to that agent
- keep all other Discord channels on my normal main agent
- do not make the critic global for all Discord
- show me the exact openclaw.json changes before applying them

The config shape is basically:

# 
{
  agents: {
    list: [
      { id: "main", workspace: "~/.openclaw/workspace-main" },
      { id: "kyaw-critic", workspace: "~/.openclaw/workspace-kyaw-critic" },
    ],
  },
  bindings: [
    {
      agentId: "kyaw-critic",
      match: {
        channel: "discord",
        accountId: "default",
        peer: { kind: "channel", id: "1493912068162846740" },
      },