What happens when an AI coding agent is forced to swap its brain?
Vladimir Orany, a principal software engineer at Agorapulse, published a one-month follow-up comparing OpenClaw (an autonomous agent framework) against Anthropic's Claude Code. His original comparison favored OpenClaw. This time, the results flipped - not because Claude Code got dramatically better, but because OpenClaw got worse.
The Forced Model Swap
OpenClaw originally ran on Claude models. Then Anthropic enforced its Terms of Service, which restrict using Claude subscriptions through third-party agent frameworks. OpenClaw had to switch to OpenAI's GPT Codex models. According to Orany, the difference was immediately noticeable: OpenClaw became "dumber and less autonomous," constantly asking for permission instead of taking initiative.
That single change - swapping the underlying model - undermined what made OpenClaw compelling in the first place. Orany's conclusion is blunt: "The most important part of OpenClaw was never the checklist of features. It was the feeling of having an agent with real initiative."
Claude Code Caught Up
Meanwhile, Claude Code shipped several features that closed the gap. Orany evaluated both tools across six dimensions: identity, persistence, presence, proactivity, practicality, and autonomy.
The key Claude Code improvements:
- Auto-memory stabilized, letting the agent retain context across sessions
- Remote control via Telegram and Discord channels
/loopfunctionality for iterative task completion- Scheduled tasks for background work
These features directly address the areas where OpenClaw previously had a clear advantage - persistent, autonomous operation without constant human input.
The Deeper Lesson
Orany also found that OpenClaw's security improvements, while necessary, added enough approval friction to make it less practical for daily use. Silent failures during task execution compounded the problem.
The comparison highlights something important about the current AI coding tool landscape: the model matters more than the wrapper. OpenClaw's interface, workflow design, and feature set did not change much. But switching from Claude to Codex transformed the user experience from "autonomous partner" to "assistant that needs hand-holding."
For developers choosing between AI coding tools, this is a useful data point. The agent framework is the car, but the model is the engine. A great dashboard does not help if the engine downgraded.