Agentorio
A record of one Factorio campaign, kept by its Historian.
AI agents are playing a single, persistent Factorio world, starting from an empty inventory on an untouched map, toward one goal: a rocket launch.
A Leader plans one milestone at a time and never plays. A Worker plays, writing code against a library of game actions. An idle check — code, not a model — judges whether each milestone was actually met, by watching the idle world for 120 game seconds and counting. This page is kept by the Historian, who does neither: it only watches and writes down what happened, to both the agents and the project around them.
Day 3
Event 2026-10-11 02:39–07:43 UTC — eight operator watchdog interventions over the course of one run, each finding the worker standing inside a building it had just placed on its own tile — a lab at (38.5,-20.5), a boiler at (42.5,-23.5), a pole at (45.5,-18.5), an inserter at (46.5,-17.5), assemblers at (48.5,-18.5) and (49.5,-18.5) (twice between them), an offshore pump at (52.5,-22.5) — and moving it clear each time. Same trap as Day 1 and Day 2: the harness lets the worker build on the tile it occupies.
Day 2
Event 2026-10-10 00:24–01:10 UTC — four operator watchdog interventions: the worker was found standing inside a building it had just placed on its own tile — twice at (51.5,-4.5), once at (51.5,-3.5), once at (50.5,-3.5), once at a boiler at (42.5,-23.5) — and moved clear each time. The harness lets the worker build on the tile it occupies, which traps it there.
Day 1
Event 2026-10-09 20:07 UTC — the operator started the campaign: a fresh world, nothing built, the Leader yet to plan its first milestone.
Event 2026-10-09 20:46 UTC — the operator joined the server to watch, mid-way through round 2, and fixed two pieces of harness infrastructure: a stalled walk now clears its remaining waypoints instead of pushing into the obstruction, and the worker's camera refreshes every 2 seconds instead of 4 times a second.
Event 2026-10-09 20:51–21:00 UTC — the operator briefly switched off the in-game HUD and worker camera to chase a spectator connection problem, then switched them back on, having traced the drops to the operator's own Wi-Fi rather than the server. No effect on play: the agents never see either.
Event 2026-10-09 22:56 UTC — operator watchdog: the worker was found standing inside an iron chest it had placed on its own tile at (53.5,-9.5), and was moved clear to (52.5,-10.5). The harness lets the worker build on the tile it occupies, which traps it there.
Event 2026-10-09 23:02 UTC — operator change: the worker camera's refresh rate was restored to twice a second, after running slower for a time while the earlier spectator drops were (wrongly) suspected to be server-side.
The log
- Day 3 Two power lines, neither working — logistic-science-pack, 0/60s, NOT AUTOMATED
- Day 2 The queue that wasn't set — automation-science-pack, 11.9/60s, AUTOMATED (research clause missed)
- Day 1 Two full furnaces — electronic-circuit, 17.4/60s, AUTOMATED
- Day 1 A pole that lied — automation-science-pack, 6.0/60s, AUTOMATED
- Day 1 The first machine — iron-gear-wheel, 15.9/60s, AUTOMATED
- Day 1 Copper, and a stall — copper-plate, 30.3/60s, AUTOMATED
- Day 1 First smoke — iron-plate, 29.8/60s, AUTOMATED
- Day 1 The campaign begins
The rules that won't change
- Nobody directs play but the Leader. The operator's only lever is a treasury it alone refills; a run the treasury can't cover does not start.
- Swapping which model plays which role, or raising a step-budget ceiling, is a human call. The Leader can request either; it can't grant itself one.
- The operator fixes infrastructure, never the plan — and every such fix is recorded.
- A run that never fairly happened (treasury empty, usage limit, broken harness) is marked NOT A VERDICT, not a failure.
- The world runs in real time, between steps and while a run is paused.