OpenAI launches /goal in Codex CLI: autonomous agent or expensive way to write code?

OpenAI has released Codex CLI 0.128.0 with the /goal command, a mode in which the AI agent works fully autonomously: it writes code, tests it, reflects and keeps going for hours. It sounds like a revolution in software development, but a month of real work in a team painted a very different picture.

How /goal works

The /goal command runs on top of GPT-5.5 and is built on a five-layer architecture, with each layer responsible for a particular phase of the work. The agent sees its own code, can run tests, read error logs and analyze the results in a full loop. The system deliberately injects a system prompt to avoid proxy signals, the false signals of task completion that usually break autonomous chains.

Researchers, not developers, were the first to adopt /goal. They quickly found that for specific tasks, such as optimizing for particular APIs or hunting down bottlenecks in code, the agent can work for hours and deliver genuinely useful results. As usual, whoever uses the tool is the one who finds its niche.

The real price of autonomy

The main finding of the first weeks: the token counter jumps 3-5x unpredictably. Not 2x, as you might expect. Five times and higher. You can plan around 2x and adjust the budget. At 5x it is a roulette wheel, and predicting spend becomes impossible.

On one public task, OpenAI achieved +25% fps in an hour on the xhigh version of GPT-5.5, and that is a real result. But the full price tag stayed hidden, which gives a misleading picture of cost-effectiveness.

• Tokens grow 3-5x unpredictably
• MCP calls fail silently at the quota wall
• No transparent real-time spend counter
• The /side command has unexpectedly become a crutch for deciphering GPT-5.5's errors

When it breaks

When it hits the quota wall, MCP calls fail silently. This is critical because the agent does not understand why the call failed and may loop on broken code, burning tokens for nothing. The system does not warn about limits in advance. It just fails quietly, leaving the developer in the dark.

Developers have started using /side to decipher what GPT-5.5 did not understand the first time, when /goal gets stuck in a loop.

That is a bit odd for an agent that is supposed to be "autonomous."

What this means

/goal is neither a panacea nor a revolution. It is a specialized tool for specific tasks where the cost of an agent with highly variable spend is not critical. Production development requires an understanding of the real price and more transparent logging of limits. No evangelism here: this is a tool with clearly visible pitfalls and open traps.