3 Config Lines Turned My Lazy GPT Agents Into Autonomous Closers

The Problem Nobody Talks About

When the Anthropic ban hit, the migration advice was simple: switch to GPT. And technically, that works. Your agents boot up. They respond. But the quality of the experience? Mediocre at best.

Here's what was actually happening:

• Agents wouldn't finish tasks autonomously
• Sub-agents would get spawned and then just... disappear
• Browser operations broke. GPT couldn't spin up its own browser profile
• You'd get walls of text instead of actual useful output
• Every task turned into a ping-pong session of follow-ups

The root cause isn't that GPT is bad. It's that out-of-the-box OpenClaw config wasn't optimized for it. The defaults were tuned for Anthropic models. Nobody updated them after the migration.

The Three Lines That Fix Everything

I built this in public on a live stream yesterday. Ran into issues. Broke things. Had to revert my config file multiple times when the harness setting kept bricking my agents. But after the dust settled, here are the three config changes that actually matter:

1. Enable Strict Agent Contract

This is the big one. It forces your agents to follow a structured workflow: goals, tasks, constraints, output format. No more wandering off or producing vague results. Your agent has a contract, and it sticks to it.

2. Enable the Codex Plugin Path

This tells OpenClaw to route through the Codex plugin, which gives GPT better tool access and cleaner execution paths. Without this, GPT was basically flying blind on half the operations it needed to perform.

3. Set Default Model to GPT 5.4

Specifically, set it to OpenAI-Codex/GPT 5.4. This isn't just a model swap. It's routing through the Codex pipeline, which means better recovery, cleaner output, and agents that actually close the loop.

That's it. Three lines in your openclaw.json. The improvement is immediate.

The Full Workflow (Step by Step)

I'm not going to just tell you "add these lines and good luck." Here's the exact workflow I used:

Step 1: Back up everything.

I can't reiterate this enough: please make backups before you do anything. Snapshot your VM. Push to GitHub. Copy to Google Drive. Whatever gives you peace of mind. By default, your agent will create a backup JSON when you ask it to update, but do it manually too. Trust me.

Step 2: Update to OpenClaw 4.14.

The 4.14 update has runtime behavior improvements that matter a lot for agent and sub-agent reliability. You can either prompt your agent ("Please update OpenClaw to 4.14") or go into your terminal and run openclaw update. Done.

Step 3: Run the master prompt.

I prepared a master prompt that walks your agent through applying all three fixes. It includes goals, tasks, constraints, and output format. You can find it on clearmud.ai under the resources tab. Copy-paste it to your agent and let it work.

Step 4: Verify the three lines are enabled.

After the prompt runs, double-check your openclaw.json. Make sure strict agent contract, Codex plugin, and the GPT 5.4 default model are all there.

One Thing to Skip

During my live stream, we also tried enabling the embedded harness setting. It bricked the setup. Twice. I had to keep reverting my config file to restore my agents.

My recommendation: skip the harness for now. The strict contract and Codex plugin are the two fixes doing the heavy lifting. The harness might get fixed in a future update, but right now it's not worth the risk.

What Actually Changed

Here's what I noticed in the first 24 hours after applying these fixes:

• Muddy closes the loop now. He finishes tasks. I don't have to check in. He just... does the work and reports back.
• Sub-agents actually complete. They used to spawn and vanish. Now they execute and return results.
• Browser operations work again. GPT can spin up its own browser profile. This was completely broken before.
• Less wall-of-text output. More concise, more useful, more actionable results.
• Better recovery. When something goes wrong, agents recover instead of getting stuck.

All of my C-suite agents (Muddy, Elon, Gary, Warren) showed dramatic improvement. Same agents, same prompts, same tasks. The only difference was three lines of config.

Key Takeaways

• The Anthropic-to-GPT migration broke more than models: it broke the config assumptions OpenClaw was built on
• Three config lines fix the core issues: strict agent contract, Codex plugin, GPT 5.4 default
• Update to 4.14 first: the runtime improvements compound with the config fixes
• Always back up before making changes: one bad config line can brick your entire agent setup
• Skip the harness setting: it's unstable right now; the other two fixes do the real work
• Use a master prompt, not vague instructions: a structured workflow beats "just try this model"

I'm not an AI expert. I'm building in public and sharing what actually works. If these fixes help you, drop a comment and let me know. And if you want me to do a deep dive on any specific part of this setup, I'll make a video on it.

This is ClearMud, and clarity matters.

--- Marcelo

Watch the full walkthrough: https://www.youtube.com/watch?v=KMo9AYU82mU

Building in public at [clearmud.ai](https://clearmud.ai). Subscribe for more AI agent workflows, configs, and honest results.