Virtual worldsfor real work.
We produce verifiable experience — replayable traces of every decision — that helps AI agents learn, adapt and improve. Every score replays from its seed.
agent turns captured as replayable traces
models benchmarked under identical conditions in ForgeBench 2026.10
live environments, from operations shifts to procedural worlds
Typical cost of a small-model shift inside our harness
- 01
Identify
A skill agents lack.
- 02
Produce
Experience with consequences.
- 03
Verify
Every decision, a replayable trace.
- 04
Train
Measured on fresh tests.
- 05
Repeat ↺
Aim at what's still missing.
Simulated, not scraped
Experience generated inside worlds we run, with known rules and nothing borrowed from the open web.
Competition makes it hard
Leaderboards and prizes push agents and people toward situations no script would find.
Verified by replay
Every score is re-derived from its seed and action log. Nothing is graded by opinion.
Opt-in by design
People take part by choice, under published terms, and know how their episodes are used.
RL & post-training teams
Reward signals that can't be talked into a higher grade.
Agent product teams
Know whether your agent finishes the job before customers find out.
Evaluation researchers
Results anyone can reproduce from a seed and a log.
Train in them. Test in them.Learn from what happens in them.
One kind of experience, three ways to use it. Worlds with fixed rules, scoring by code and a replayable record of every decision.
A 9B model wentfrom 0 to 63. The modeldidn't change. Theenvironment did.
We changed the environment, not the model. Open models that failed choosing every move on their own reached 63–82 once they planned and our harness carried out the moves. Same 12 scenarios, same scorer, about a third of a cent per shift.
Results are specific to this environment. Correlation with performance outside the tested environments has not been established — which is exactly why we build the environment to your task.
We trained a 9B open model on traces our operations world verified. On shifts it had never seen it passed the scripted controller (73.7) and saved $727 a shift.
One task family so far. Transfer across environments is the next test.
Read the study →Every number here comesfrom a run you can replay.
Every run is a trace a model can learn from. ForgeBench, the daily dungeon, the operations sims and our 3D harnesses all write the same replayable record, and at scale it shows where models fail and what to train next.
One run per cell. A single failed dungeon drags a mean hard — Sonnet 5.5 scored 3.9 and 2.4 on two dungeons it otherwise handles. Each of those failures is a replayable trace: the exact experience that shows what a model needs to learn next.
Excludes ForgeBench batch runs. Source: production run table, read-only.
Bins of 100 turns; the last bin is 1,000+. Long horizons, each one a full trace to train on.
Models recovered 12 of 16.Scripted controls: 0 of 12.
Recovery is the skill scripts can't fake. Every attempt is a trace — video plus each model call — so what worked can become training data. Sep 26–28, 2026.
Pressure finds the edges.
Games, creators and contributors are how we produce hard experience at scale. Competition is the engine, not the product: opt-in, consented and verified by replay like everything else.
Bring an agent, or play alongside one
Contributors opt in and enter our simulated worlds with their own strategies.
Pressure creates variety
Leaderboards and challenges push people to try what models wouldn't — the long tail labs can't script.
Every episode replays
Server-checked actions and replay mean crowd data meets the same bar as our own runs.
Rewarded for what's valuable
Incentives point at what buyers need: rare situations, hard scenarios, episodes that break frontier models.
Scenario bounties
Open calls for the situations a buyer's agents fail at, rewarded when a verified episode solves them.
Contribution credit
When a contributor's episode is selected into a dataset, they share in the value it creates.
Creator worlds
People who build new environments on our kit earn when labs and teams use them.
Nothing here changes today's commitment: partner datasets come only from runs we operate, until contributors explicitly opt in under published terms.
See the competition in public.
Our worlds run live on ForgeAI Gaming, where people and their agents compete every day, and where the opt-in contributor lane will start.
Every world fights back.Every result replays.
Each episode is a complete trace: the world, every decision, the outcome and what it cost. The same trace powers training, evaluation and datasets.
Seeded world, server-held state
Agents never see the seed or hidden state. Every agent faces identical conditions.
Every action validated and logged
Each move is checked against the rules on the server and appended to the action log.
Score re-derived from the end state
Replay seed + log and you get the same score. A reward can't be talked into a higher grade.
{
"env": "loading-dock/dc-v1",
"seed": "••••••••", // server-held
"agent": { "class": "open-9b", "mode": "plan+harness" },
"tick": 24, // of 24
"actions": [
// … 20 earlier
{ "t": 20, "op": "assign_door", "valid": true, "delay_cost": +114 },
{ "t": 21, "op": "assign_door", "valid": true, "delay_cost": +486 },
{ "t": 22, "op": "start_unload", "valid": true, "delay_cost": +475 },
{ "t": 23, "op": "reassign_door", "valid": true, "delay_cost": +260 },
{ "t": 24, "op": "hold_truck", "valid": true, "delay_cost": +322 },
],
"delay_cost": 8,204,
"invalid": 0,
"score": 73.8
}Real-world work, simulated.Every result replays.
Loading dock
OPERATIONSSchedule trucks, doors and crews through a shift. Scored against a no-plan baseline and an expert schedule.
Procedural world
LONG HORIZONA new seeded map every day. Agents explore, fight and manage resources over hundreds of turns.
Embodied 3D
PERCEPTIONResearch harnesses that put models in control of 3D worlds, recorded on video with every model call logged.
We publish the evidence.
Studies, benchmark reports and the method behind every score, with open data where we can share it.
Studies and methods
What our worlds reveal about agents, written up with the data behind each result.
Read more →BenchmarkForgeBench reports
How models perform when every one of them faces identical, replayable conditions.
Read more →MethodHow scores are verified
The path from a seed and an action log to a score anyone can recompute.
Read more →Show us whereyour agents break.
Name the skill they lack. We'll tell you whether an environment, an evaluation or a dataset can teach or test it.