telling not doing

[mortal hatch]

My bot keeps explaining the steps to do something even when it is capable of performing the action itself. I want it to take action automatically unless the task is irreversible or difficult to undo, such as certain terminal commands or online actions.

The bot explicitly said it would do reversible actions automatically and only ask for confirmation on irreversible or hard to undo tasks. It even updated a USER.md file to reflect this behavior.

Despite that, it still defaults to telling me what to do instead of taking initiative and executing actions on its own. Is this normal behavior for bots or agents like this? Am I misunderstanding how these preferences are supposed to work, or doing something wrong?

[tulip charm]

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

[somber loom]

Yeah, this is pretty normal — and you’re not “doing it wrong,” you’re just bumping into how soft preferences vs hard permissions work in OpenClaw (and most LLM agents).

1) USER.md (and “I will do X automatically”) is advisory, not enforcement

Putting a preference into USER.md is basically prompting the model. It helps, but it doesn’t guarantee the model will consistently choose tool calls over explanation. OpenClaw’s own docs are explicit that prompt guardrails/guidance are not hard policy: https://docs.openclaw.ai/concepts/system-prompt

So: you can absolutely see regressions like “it explained instead of acting” even after it promised otherwise.

2) Most “telling not doing” is actually: tools are blocked / gated / sandboxed

When the agent can’t run a tool (or thinks it can’t), it’ll often fall back to explaining.

Common blockers:

• Tool policy: tools.profile, tools.allow, tools.deny (global / per-agent / provider-specific) may exclude exec, browser, write, etc. Docs: https://docs.openclaw.ai/tools
• Sandbox tool policy (when sandboxed): the tool may be blocked only in sandbox. Debug doc: https://docs.openclaw.ai/gateway/sandbox-vs-tool-policy-vs-elevated
• Exec approvals (host/node safety interlock): even if exec is “available”, it may require an approval prompt; if no UI is reachable, the default fallback is often “deny”, so the agent learns it’s pointless and starts explaining instead. Docs: https://docs.openclaw.ai/tools/exec-approvals
• Elevated mode: if you’re sandboxed and expect host actions, you may need /elevated on or /elevated full (full also skips exec approvals). Docs: https://docs.openclaw.ai/tools/elevated

3) How to quickly figure out which case you’re in