Set the objective.
Declare the goal, available tools, success conditions, safety constraints, and starting world state.
Archetype · World runtime
Build Worlds.
A runtime for persistent, branchable environments where agents act, systems evolve, and every possible future has somewhere to begin.
Selected world
Earth System
Identity
w_0001
State
Persistent · live
Lineage
Root · 04 branches
World artwork: NASA / JPL-Caltech
Query Service · Projection 01
View World.
One world state, rendered many ways. The Query Service feeds every view and projection without changing the world beneath it.
query world("w_0001") {
projection(kind: VISIBLE, tick: 2409841)
}
Projection
Visible spectrum
Source
Query Service
World tick
2,409,841
State hash
b54f…9a20
Earth data: NASA Goddard / SVS
Lineage · Counterfactual runtime
Fork World.
Branch from any tick. Change a condition. Preserve the parent world and compare what follows—from an ocean planet to an Earth without water.
Root world · tick 2,409,841
Counterfactual futures
03 / Lineage
Selected branch
Temperate baseline
Identity
w_0001
Divergence
Root state
Outcome
Stable
Autonomous work · Observable state
Agent Missions.
Give an agent a goal, tools, constraints, and a world. Archetype records every action and state transition so a mission can be understood—not merely scored.
Let agents act.
Agents perceive projections, operate on entities, and change a persistent environment one auditable step at a time.
See what changed.
Query traces, compare branches, replay decisions, and render the same outcome through the views that matter.
Repeatable reality · Counterfactual proof
Physical AI Evals.
Evaluate autonomy inside dynamic, stateful environments. Replay the same mission, fork the moment that matters, and measure whether an agent can recover when the world changes.
Return to worlds ↑
Environment
Versioned worlds
Observation
Queryable projections
Comparison
Counterfactual forks
Evidence
Signed state history
Open source · Run it yourself
Try Archetype.
No account. No sales call.
Give Archetype a mission and build the smallest working world. Each prompt inspects the source, verifies the interfaces, and starts with a credential-free smoke test.
Run a mission.
Create a persistent world, give an agent an objective, record its actions, and query the resulting state.
Evaluate a policy.
Run VLA-JEPA against a LIBERO task, preserve the trajectory as world state, then fork the starting conditions for comparison.