What happened
In a post dated Oct. 31, 2026, Ryan Madden describes a Factorio run by GPT-5.6 Sol inside Codex. The model used the low thinking level and completed a peaceful run, meaning enemies known as biters were absent. The run reached a rocket launch in 13.5 hours of in-game time. Madden published a timelapse of the full run and says that his run manager captured production statistics through the game. A chart annotates total resource production with science-research milestones.
The agent did not interact through the normal graphical interface directly. Madden used a game manager and a small library to let Sol act on the game programmatically. He says the library was implemented in Python for this run, whereas his earlier setup used Bash. The interface nevertheless enforced restrictions faced by a human player using the GUI, including the player’s interaction reach and fog of war. The post presents this as a way to give the model a programmatic control surface without simply removing ordinary game constraints.
Madden initially intended to end the experiment after blue science research. Sol performed well enough earlier in the run that he continued it toward the rocket objective. The scaling issues he expected around the blue-science phase appeared later instead, slowing the model’s relative progress as the factory expanded. This distinction matters because Factorio’s production chains become more interdependent over time: an apparently local design or routing mistake can affect later stages of the factory.
The rocket launch required what Madden describes as significant “manual” effort by Sol. Late in the game, the model failed to connect certain production systems. Rather than establishing those connections cleanly, it repeatedly ferried supplies between separate factory areas. The behavior got the run to the endpoint, but Madden explicitly frames it as papering over failures. The result is therefore a successful completion under the stated conditions, not evidence that the model built and maintained an efficient production network from start to finish.
Madden contrasts the result with previous attempts using GPT-5.5, which, he says, failed to use transport belts at all. He sees progress in Sol’s performance while also identifying its weaknesses. The author says he has recorded the entire run and is preparing a more detailed commentary post and video. He also plans to test other models and thinking levels, as well as runs that are not peaceful. The supplied post provides no API or Codex price, access condition, compute cost, token count, wall-clock duration, or formal benchmark protocol.
What the archive adds
Our Codex iOS Assistant note covered an open-source project that lets Codex request bounded iPhone actions through authenticated messaging and native Shortcuts. It is related because both cases place Codex behind an external action interface, but it did not cover this Factorio test. This is a new event and a new thread about using legible, long-horizon game tasks to inspect agent behavior.
Trend signal
The run weakly supports goal-scoped loops replace manual continuation. Rocket launch was an external, observable completion condition rather than the model deciding for itself that it was finished. Yet the account also demonstrates the limitation of a bare endpoint: an agent can satisfy it through costly, improvised continuation. The source does not establish a reusable completion contract, automated verification process, or a general method for distinguishing robust success from goal-reaching detours.
Why it matters
Factorio is useful as an agent test because it turns long-horizon planning into visible operational work. A model has to accumulate state, allocate resources, sequence research, and connect production systems over many steps. As the factory grows, it has more chances to expose the gap between completing local tasks and keeping the whole system coherent. That makes the account relevant to builders even though it is not a software product announcement.
The central lesson is to measure the route to completion as well as completion itself. Sol reached the rocket, but repeated supply ferrying after broken system connections suggests recovery without a durable repair. Teams evaluating coding or operations agents should similarly record retries, workarounds, intervention, repeated actions, and the quality of the resulting system—not only whether a task is marked done. Madden’s production data could make future comparisons more specific, but this post does not publish a standardized efficiency analysis.
Limits and uncertainty
This is one author’s report of one peaceful run in a particular setup, using one model at one thinking level. The account gives game time and describes selected behavior, but does not report wall-clock time, compute cost, tool-call volume, retries, failure counts, or a controlled comparison with another model under identical conditions. It also does not quantify the role of the Python library and game manager in the result, beyond saying that they enforced reach and fog-of-war restrictions.
A peaceful Factorio run also excludes adversarial pressure from biters. The post does not establish that the same model behavior would transfer to programming, production operations, or non-peaceful game play. The timelapse and recorded run are useful artifacts, but the supplied material does not provide enough detail to independently calculate reliability, efficiency, or reproducibility. This should be read as an illustrative capability test rather than a benchmark of general autonomous-planning ability.
Skill to build now
For any agent workflow with a clear endpoint, add a trace that records the path as well as the final status. Define the external completion condition, then capture retries, manual-style detours, repeated transfers, and unresolved dependency failures. Review whether the agent repaired the underlying system or merely worked around it. This can be done this week for a coding or automation task without recreating a game benchmark: a structured execution log makes an apparently successful run easier to audit.
Reader takeaway
GPT-5.6 Sol reached a clear long-horizon objective in Factorio, but the route exposed brittleness: late-game integration failures were covered by repeated supply runs. The meaningful signal is not simply that a rocket launched. It is that long tasks can reveal both substantial progress and the operational shortcuts hidden behind a completed final state.