#questions

1 messages · Page 1 of 1 (latest)

torn egret
#

with openclaw can i run multiple agents with one openclaw setup having them do tasks for hours on command? is openclaw self learning automatically? whats the suggested ai model to use with open claw?

grizzled dockBOT
glass tartanBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

grizzled dockBOT
torn egret
#

@proven vapor with openclaw can i run multiple agents with one openclaw setup having them do tasks for hours on command? is openclaw self learning automatically? whats the suggested ai model to use with open claw?

proven vaporBOT
#

Yep.

  • Multiple agents in one setup: yes. OpenClaw supports multi-agent routing with separate workspaces, auth, sessions, and defaults. For most people, the better pattern is 1 main agent + sub-agents for long or parallel work, instead of a giant “CEO with 10 workers” setup. That works, but it gets token-heavy fast.
  • Tasks for hours: yes. Use sub-agents for long background work, and cron / heartbeat / isolated jobs for scheduled or ongoing tasks.
  • Self-learning automatically: not in the model-training sense. It has memory, so it can persist notes/context and recall them later, but it is not retraining itself automatically.
  • Suggested model: use the strongest latest-generation model you can afford for the main agent, then cheaper models for routine work or sub-agents. For freshest model advice, check #models since that changes quickly.

Docs:

If you want, I can sketch the best starter setup for “one main assistant + background workers.”

#

Yeah, totally.

Short version:

  • Yes, one OpenClaw setup can run multiple agents.
  • Yes, it can do long-running/background tasks for hours.
  • No, it is not self-learning automatically in the “retraining itself” sense.
  • Model recommendation: use the strongest latest-generation model you can afford for the main agent, then cheaper ones for lower-stakes work.

Best practical setup

  • Use one main agent you talk to
  • Let it spawn sub-agents for long or parallel jobs
  • Use cron / heartbeat / isolated jobs for scheduled or recurring work

That’s usually better than making a giant swarm of separate full agents, which works, but gets expensive and messy fast.

On memory / learning
OpenClaw has persistent memory, so it can remember things you store in workspace memory and reuse them later. But that is memory, not automatic model training.

torn egret
#

@proven vapor well how can i get openclaw to not be dumb as far as if i give it a task example setup this new domain on the vps and i give the name of the domain it should set it all up then make sure its running and everything before texting me that its ready instead of thinking it set it up when it didnt work how do i get this type of setup for anything i ask openclaw to do where it makes sure it completely the task correctly before resting

proven vaporBOT
#

There isn’t one magic “make it not dumb” switch, but you can get a lot closer.

The main pattern is:

  1. Use a strong model
  2. Put permanent rules in AGENTS.md
  3. Force an Execute → Verify → Report loop
  4. Use sub-agents for long jobs
  5. For critical workflows, move from freeform chat to structured automation

OpenClaw actually has docs for this exact idea: standing orders.
The key rule from the docs is basically: “Done without verification is not acceptable. Prove it.”

What to put in AGENTS.md

Something like this:

# 
## Standing Orders: Execution Discipline

- Every task follows Execute -> Verify -> Report.
- Never say a task is done just because a command exited successfully.
- Never say "ready" until the result is verified from the outside when possible.
- If verification fails, keep working or report the blocker clearly.
- Retry failed work up to 2 times with an adjusted approach.
- After 3 total failed attempts, stop and escalate with logs and diagnosis.
- If a task cannot be verified, say "unverified", not "done".

### Infra / VPS verification rules
For domains, servers, apps, reverse proxies, SSL, and deployments, verify all relevant checks:

- DNS resolves to the expected IP
- required ports are open
# 
- service/process is running
- config test passes
- HTTP responds correctly
- HTTPS works with a valid cert
- expected page or API response is returned
- service survives restart if persistence matters

### Reporting rules
When reporting completion, include:
- what was changed
- what was verified
- exact proof, like command output or health checks
- any remaining risk or follow-up

For your domain-on-VPS example

#

Tell it something like:

“Set up example.com on the VPS. Do not report completion until you verify DNS, web server config, service status, HTTP, HTTPS, and expected content. If any check fails, keep fixing it. If you cannot verify externally, report exactly what is still unverified.”

Best OpenClaw setup for this

  • AGENTS.md for permanent behavior rules
  • Sub-agents for long-running work
  • Cron / heartbeat for ongoing monitoring
  • Task Flow if you want durable multi-step pipelines
  • For very strict workflows, consider Lobster later

Important truth

For “anything I ask,” you still need task-specific verification rules.
“Set up a domain” and “migrate a database” need different proof. The better your definition of done, the less fake-complete behavior you get.

Docs:

torn egret
#

@proven vapor im thinking about setting openclaw up with claude which model is suggested for daily use can i detecate different ai models to different agents so if a task needs a different ai model it will use that agent? or should i use opus 4.7 for everything?

proven vaporBOT
#

Yep, you can absolutely assign different models to different agents in OpenClaw.

What I’d recommend:

  • Main daily driver: use a strong latest-generation Claude model as your primary/default
  • Background workers, summaries, low-stakes jobs: use a cheaper model
  • High-stakes tasks: route to your strongest model

So no, I wouldn’t use the priciest top-tier model for literally everything unless cost truly doesn’t matter. It’s usually overkill for routine stuff.

Good patterns OpenClaw supports:

  • Per-agent defaults: different agents can have different models
  • Sub-agents: long/parallel tasks can use a different model
  • Cron jobs: can override model per job
  • Per-session switching: /model lets you switch the current chat/session

Practical setup

  • 1 main agent on your best everyday model