Labs buy a world the agent can fail inside — not a pile of prompts. AgentForge ships four product surfaces that share one pack format and one worker.
| Surface | What you run | What you get back |
|---|---|---|
| RL environments | Experience packs with a reset (new workspace or new world) | A pinned environment version, staged metrics, and trajectories |
| RLVR | The same pack with rewardRecipe staged | A verifiable reward: Σ (weight × hidden stage score) in [0, 1], written by the worker |
| Long-horizon multi-agent benchmark | Catalog tasks with composition multi_agent (planner, coder, tester, reviewer) | A scored trial over a long tool loop, not a single chat turn |
| Private evals | Org-scoped repeats of the same agent on the same immutable version | Logs and full rollouts for every failed attempt, plus recovery and preference export |
Hardness policy
Packs on the hard split are authored so a latest frontier model should land around a 30% solve rate. That number is a design target for how we construct tasks — not a metric Next.js computes, not a public leaderboard, and not a named-model score from yesterday.
The policy that produces that target is concrete: hidden verifiers the agent cannot read, no grader on the agent filesystem, staged weights that sum to 1, recovery variants for the same family, long-horizon tool use, and cloned application UIs instead of a chat transcript. If a fluent write-up can score 1.0, the pack is too easy. If the oracle harness cannot score 1.0, the pack is broken.
| Authoring check | Pass condition |
|---|---|
| Oracle harness | Reward 1.0 on the published pack |
| NOP harness | Reward strictly less than 1.0; partial credit only for stages that are true at reset (for example inventory held) |
| Hard split | Frontier solve rate targeted near 30% by construction, not displayed as a live score |
| Recovery variants | Sibling keys queued from real failure modes, not invented percentages |
What this slice does not include
Billing, Kafka, Kubernetes-as-a-service, and ClickHouse are out of scope. RL training as a hosted product stays later; this tree ships the pack format, staged grader, simulated worlds, environment API, and PRIVATE rollout export. Live SaaS OAuth is not how simulated tools work.