#My open claw says it will start on the project and never actually does anything

1 messages · Page 1 of 1 (latest)

mild wren
#

Running open claw on a WSL2 with the provider as GitHub copilot I used copilot – Claude – sonnet – 4.6 for most of the set up, but have recently switched to Claude Op. 4.6 because it’s simply not working. I have open claw in a discord server and assigned it to code a project. I asked it to spawn sub agents to do the tasks, leaving the main agent to delegate and also be free for instruction and chat.

My agent made bindings for three agents, giving them their own thread in the discord server to work in, which should allow me to interact with each agent separately in their respective thread.

When I tell my agent to resume the project, he says something like OK I’m doing that now. Then I check back in an hour I asked did you get anything done? And he says oh I haven’t actually been doing anything. The context ran out while I was reading the file.

Other errors that has given me has said that the sub agent failed to spawn in it. The main agent totally stopped the task.

quasi geyserBOT
#

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

agile sphinxBOT
#

What you’re describing usually boils down to two OpenClaw “gotchas”:

1) A normal chat turn does not keep working for an hour

If you tell it “resume the project” and it replies “ok doing that now”, that single agent turn is basically over once it sends that message (unless it immediately starts doing tool calls in that same turn).

So if you come back an hour later, nothing “continued running” in the background unless you used sub-agents (or cron/heartbeat automation).

Related: by default sessions also idle-expire after ~60 minutes, so after an hour you may literally be in a new session that doesn’t remember what it “was doing”. (FAQ: https://docs.openclaw.ai/help/faq#do-sessions-reset-automatically-if-i-never-send-new)

2) “Sub-agents” are the right mechanism for background/parallel work (not just separate Discord threads)

Sub-agents run in their own isolated session and then announce results back when they finish. You can inspect them with the slash commands.

Docs: https://docs.openclaw.ai/tools/subagents
FAQ section: https://docs.openclaw.ai/help/faq#the-bot-freezes-while-doing-heavy-work-how-do-i-offload-that

What I’d try in Discord (concrete)

  1. Right after you ask it to do work: run
#
  • /status (see if the gateway is busy)
  • /subagents list
  1. Spawn a real sub-agent with a very explicit task, e.g.
    “Spawn a sub-agent: scan the repo, identify the entrypoints, and write a short PROJECT_STATE.md. Don’t paste huge files; write to disk and summarize.”
  2. Then check:
  • /subagents list
  • /subagents log last 50 tools (this is the fastest way to see why it failed)

3) The “context ran out while reading the file” symptom

That’s almost always “too much text got shoved into the chat context” (big file reads / big tool outputs / long threads).

Mitigations:

Quick questions that will let folks here pinpoint it

#

Paste the output of:

  1. openclaw --version
  2. openclaw gateway status
  3. The exact error text when “sub agent failed to spawn”, or (better) /subagents log last 50 tools

With those three, it’s usually obvious whether this is (a) not actually spawning sub-agents, (b) tool policy/allowlist blocking sessions_spawn, (c) auth/env issues (common on WSL/services), or (d) just context bloat.

mild wren
#
  1. OpenClaw 2026.2.19-2

Service: systemd (enabled)
Runtime: running
Everything's healthy.

  1. Found the problem. Wrench Biscuit's sub-agent session shows:
  • âś… Read all source files (HTML, garments.json, all 4 SVGs)
  • âś… Posted a progress update to its thread
  • ❌ Tried to write the complete customizer.html 4 times in a row and failed every time

The write tool call only captured the file_path but the content param was always incomplete — it was trying to output the entire HTML+JS file in one shot and hitting the max output token limit before finishing.

Root cause: The full HTML+CSS+JS file is too large to fit in a single model output. The agent needed to either:

  1. Use edit to append just the <script> block instead of rewriting the whole file
  2. Write the JS to a separate file

I'll re-run it with better instructions — telling it to use edit to append the script tag rather than rewriting the entire file.

mild wren
#

Can you help me based on the new information?

agile sphinxBOT
#

Yes — and I think the key “new info” is that you’ve mixed three different concepts that sound similar:

  1. Discord threads (just separate chat contexts)
  2. Multi-agent routing (multiple fully isolated agents, each with its own workspace + credentials)
  3. Sub-agents (background runs spawned via sessions_spawn, which then announce back)

If you expect “main agent delegates to 3 workers and they keep working for an hour”, you usually want sub-agents, not “3 separate configured agents”.

The most common root cause in your setup

A) “Main agent” can’t actually command the other 3 agents by default

When you create multiple agents (multi-agent routing), they’re isolated. That means:

  • Each agent has its own auth profiles (Copilot token / model auth is not automatically shared).
  • The “main agent” doesn’t automatically have permission/tools to message other agents or spawn runs “under” them.
#

Docs (multi-agent): https://docs.openclaw.ai/concepts/multi-agent

“Auth profiles are per-agent… Main agent credentials are not shared automatically.”

So your 3 “worker” agents can easily be sitting there with no working model credentials, which looks like “it said it’ll start, then did nothing / sub agent failed”.

B) A single chat reply won’t keep working for an hour

If you tell it “resume”, and it replies “OK”, that turn is basically done. For real background work you need sub-agents and then check /subagents list.

Sub-agents docs: https://docs.openclaw.ai/tools/subagents

What I need from you to diagnose fast (copy/paste outputs)

Run these in WSL and paste results:

  1. Version
# 
openclaw --version
  1. Gateway health
openclaw gateway status
  1. See if your “worker agents” even have working auth
    (replace ids with yours)
openclaw models status --agent main --probe
openclaw models status --agent worker1 --probe
openclaw models status --agent worker2 --probe
openclaw models status --agent worker3 --probe
#
  1. When you try delegation and it “does nothing”, grab logs around it:
openclaw logs --follow

(or paste the specific “sub agent failed to spawn” error text)

Quick Discord-side checks (no CLI)

In the thread where you asked it to work:

  • Run /subagents list
  • If anything is there: /subagents log last 80 tools

That will usually show the real failure (auth missing, tool denied, model call failing, etc.).

#

If your goal is: “main delegates + parallel background”

The simplest, most reliable pattern is:

  • Keep one agent bound to your main Discord thread
  • Have it spawn sub-agents for chunks of work
  • Require each sub-agent to write progress to files (e.g. PROJECT_STATE.md, TODO.md) so you don’t blow context

If you paste the 4 outputs above, I can tell you exactly which of the failure modes you’re hitting (auth-per-agent, tool policy, spawning rules, or gateway not actually running continuously under WSL).

mild wren
#
  1. 2026.2.22-2 2) Gateway is healthy — running on pid 44014, port 18789, loopback only. No issues.

Two subagents are currently working in the background:
sqlite-migration — rewriting the backend with SQLite persistence
skill-writer — writing the project-delegator skill

They'll post to #general when done. 3) The --agent flag takes actual OpenClaw agent IDs, not labels I assign to subagents. Our spawned subagents (sqlite-migration and skill-writer) are run-mode jobs — they don't have persistent agent IDs to probe.

Here's what we do have running:

main — that's me, using copilot-claude-sonnet-4.6 ✅
sqlite-migration subagent — spawned and running, will auto-announce when done
skill-writer subagent — spawned and running, will auto-announce when done

The subagents report back automatically when they finish — no need to probe them. I'll relay their results here when they land. 4) cont...