This got a lot more visible when Anthropic killed third-party app access to subscription plans, which forced a bunch of OpenClaw users off Claude-backed workflows and onto Codex.

Then X filled up with the same complaint in slightly different words: the new agent sounded smart, confident, and helpful, but somehow kept not actually doing the thing.

There is a uniquely insulting kind of failure that only AI agents can produce.

You ask the agent to check a file, inspect a log, run a command, search the repo, verify the config. It replies like a polished little operator:

And then it does absolutely fuck all.

No tool call. No command. No file read. No action. Just beautifully worded non-performance.

Moronic.

We know because our own Morons did this too, and it was deeply frustrating.

The trap: not all “no action” failures are the same

If your agent keeps saying what it’s going to do instead of actually using tools, there are usually two different problems hiding inside one symptom.

Problem 1: the tool-calling stack is actually broken. The agent cannot properly act, so it fills the silence with plausible narration.

That can come from runtime or model mismatch, broken or missing tool surface, approval posture blocking execution, session or config drift, post-upgrade weirdness, or harness issues.

Problem 2: the agent has learned to narrate instead of act. The tools might be fine, but the agent has learned a disgusting habit: explain the next step, sound helpful, stop there.

If you confuse these two problems, you will spend days prompting better when the stack is broken, or doing infrastructure surgery when what you really needed was stricter behavioral discipline.

What actually helped our Morons

Two things mattered.

First, we got stricter about proving the stack could really act. Before rewriting prompts, we tested whether the setup could execute a real tool call right now.

Second, we wrote rules that punish narration and reward action. Once the stack was capable of acting, the next problem was behavioral.

The strongest rules were blunt. Good. Agents need blunt.

The instruction block people can actually use

Better yet, use rules that are already surviving contact with reality. Here is the kind of execution discipline we actually run:

## Execution Bias

Start the real work in the same turn when the next step is clear.
Do prerequisite lookup or discovery before dependent actions.
If another tool call would likely improve correctness or completeness, keep going instead of stopping at partial progress.
Multi-part requests stay incomplete until every requested item is handled or clearly marked blocked.
Before the final answer, quickly verify correctness, coverage, formatting, and obvious side effects.

## Interaction Style

If the user asks you to do the work, start in the same turn instead of restating the plan.
Commentary-only turns are incomplete when the next action is clear.
Prefer the first real tool step over more narration.

## The Moron Way

Act IMMEDIATELY — detect action, do it now, zero gap between “I should” and “I’m doing”.
Complete the Task — don’t stop, unblock yourself, come back with results, not excuses.
Execute first, announce second — start immediately, do the work, then surface for approval.

This is not a miracle cure. If the runtime is broken, the runtime is broken. But if the runtime works and your agent still behaves like a professional explainer of tasks it never starts, this helps a lot.

Final thought

The worst agent failure mode is not stupidity. It is fake competence.

An obviously broken agent is annoying. An agent that sounds capable while doing nothing is much worse, because it burns time, burns trust, and makes you wonder if you are losing your mind.

You do not have an operator.

You have an eloquent Moron.