Someone in the community ran /goal for 21 hours straight, burning 900 million tokens — OpenAI has transformed the AI coding assistant from a "chat tool" into an "unattended worker."
What this is
/goal is a command added by OpenAI on April 30 in Codex CLI (command-line coding tool) version 0.128.0. It's not an ordinary prompt, but a goal lifecycle management mechanism: you give a goal, and Codex pushes forward round by round until completion, manual pause, or hitting the token limit.
A four-layer architecture supports this mechanism: the persistence layer stores goals as independent states, so compressing conversation history won't lose the goal; the interface layer allows clients to read and write states; the model tool layer only permits the model to query and declare completion, not to pause or tamper on its own — this is the safety boundary; the runtime continuation layer automatically injects prompts during each idle round to drive the next step.
The most notable design is "completion audit." Before the model determines a goal is achieved, it must map the goal to a concrete deliverables list and verify each item with real evidence. It cannot rely on auxiliary signals like "tests passed" or "code written." This targets the model's tendency in long tasks to claim completion early and then slack off.
The token budget is a soft stop: hitting the limit won't cause a crude interruption, but instead lets the model wrap up, summarize progress, and list remaining work, so you get a readable report when you open the terminal the next day.
Industry view
Positive voices see this as a critical step for AI coding from "assistant" to "autonomous worker." Previously, using AI to write code required constant supervision; now you set a goal and validate the next day. For repetitive tasks like batch refactoring, adding tests, and fixing bugs, the efficiency gain is clear. OpenAI President Greg Brockman calls it "built-in Ralph Loop++" (Ralph Loop refers to a workflow using scripts to repeatedly drive an agent toward the same goal); the built-in version supports cross-session recovery without rebuilding context each round, yielding more stable output quality.
But we note three risks: First, cost overruns. 21 hours burning 900 million tokens — if the goal is poorly set or more complex than expected, the bill gets ugly. Although there's a token budget limit, it's only a soft stop; the preceding consumption is irrecoverable. Second, drift. Issue #19910 reports: when context compression is triggered mid-call, the continuation prompt is not re-injected, and subsequent rounds may lose the goal. Stability for ultra-long tasks still needs observation. Third, limited scenarios. /goal suits repetitive, listable batch tasks; for exploratory tasks requiring frequent human judgment, using it wastes tokens and easily gets stuck.
Impact on regular people
For enterprise IT: Grunge work like code migration and tech debt cleanup can use /goal to "set goal — validate next day," but you must first establish token budget controls and goal validation processes, otherwise the cost exposure is too large.
For individual careers: Programmers' work focus shifts further from "writing code" to "defining goals and validating results." Those who can clearly decompose tasks and write auditable goals are better suited for this type of tool than those who only know how to write code.
For the consumer market: No direct impact yet. /goal is a developer tool, but its design concepts — goal persistence, completion audit, budget control — are likely to be borrowed by other AI products and appear in automation scenarios for general users.