OpenAI launched /goal in Codex CLI: autonomous agent or an expensive way to code?

OpenAI released Codex CLI 0.128.0 with the /goal command — a mode where an AI agent works completely autonomously: writes code, runs tests, reflects, and continues working for hours. It sounds like a revolution in development, but a month of real-world team experience revealed a quite different picture.

How /goal Works

The /goal command runs on top of GPT-5.5 and is built on a five-layer architecture where each layer handles a specific work phase. The agent sees its own code, can run tests, read error logs, and analyzes results in a complete cycle. The system specifically injects the system prompt to avoid proxy signals — false completion signals that typically break autonomous chains.

It wasn't developers who first adopted /goal, but researchers. They quickly discovered that for specific tasks — like optimization for particular APIs or finding bottlenecks in code — the agent can work for hours and deliver genuinely useful results. As it usually happens: whoever uses it finds their niche.

The Real Price of Autonomy

The main discovery of the first weeks: token count jumps 3-5x unpredictably. Not twofold, as one might assume. Five times and higher. Twofold can be planned and budgeted for. Fivefold — that's already a lottery, and predicting consumption becomes impossible. On one of OpenAI's public tasks, they managed to achieve +25% fps in an hour on the xhigh version of GPT-5.5 — that's a real result. But the full cost was kept hidden, creating a misleading impression of profitability.

• Tokens increase 3-5x unpredictably
• MCP calls silently fail at quota wall
• No transparent real-time expense counter
• /side command unexpectedly became a workaround for decoding GPT-5.5 errors

When It Breaks

When hitting the quota wall, MCP calls silently fail. This is critical because the agent doesn't understand the reason for the failure and may repeatedly cycle through non-working code, wasting tokens. The system doesn't warn about limits in advance — it just silently crashes, leaving the developer in the dark.

Developers began using /side to decode what GPT-5.5 didn't understand on the first attempt when /goal gets stuck in a loop.

It's somehow odd for an agent that's supposed to be "autonomous."

What This Means

/goal is not a panacea or a revolution, but a specialized tool for specific tasks where the agent's high variable cost is not critical. For production development, it requires understanding of the real price and more transparent limit logging. No evangelism: it's a tool with clear gotchas and open pitfalls.