RL + multi-autonomous-agents roadmap
Status: parked — revisit later (after the distillation thread closes and a first GRPO loop exists). Captured so we can pick it up cold.
The realization
The fleet’s AI game (a bunch of characters living in a world autonomously) is a multi-agent RL environment. Environments are the scarcest ingredient in RL — Prime Intellect built a 2,500-environment Hub precisely because that’s the bottleneck. We already have a live one. That’s a real, ownable asset.
What RL concretely does here
A game produces the RL loop for free: character observes world → acts → world updates → outcome = a trajectory. RL needs three things, all present:
| RL ingredient | In the game |
|---|---|
| Policy | the character’s model brain — the distilled small local model (cost-compression lever) |
| Environment | the world |
| Reward | a measurable in-game outcome (goal reached, need satisfied, task done, survived, won an interaction) — clean/verifiable = RLVR, the tractable kind |
Wire them together → self-improving NPCs: log trajectories (B22 recorder:
observation → action → reward per turn), run GRPO on the policy, and the
character gets better at living in the world over time — trained on a Mac, in
our own environment.
First experiment (the PoC to run when we return)
- Pick one measurable behavior (a character completing a specific goal / winning a negotiation).
- Instrument the game to log each turn
(observation → action → outcome→reward). - Run GRPO on a small local model (the distilled specialist) as that character’s brain; reward = the in-game outcome.
- Measure: does that character’s success rate climb over RL iterations vs the base brain?
A “yes” = a self-improving agent, trained on a Mac, in our own world — the “after autonomous agents” thesis demonstrated, not theorized.
Staged roadmap (multi-autonomous-agents)
- Stage 0 (have): autonomous characters = model + loop + world.
- Stage 1: instrument trajectories + rewards (B22) → the game becomes an RL environment.
- Stage 2: single-character self-improvement (GRPO on in-game reward) → self-improving NPC.
- Stage 3: multi-agent dynamics — characters co-adapt, specialize, learn from each other; emergent social behavior (generative-agents / AI-town, but local + RL-trained, not just prompted).
- Stage 4: the game as a general RL gym — agents trained in-game transfer out to other fleet projects; the eval/judgment layer verifies them.
Why this fits the north-star
It fuses everything into one Mac-runnable system, on an asset we already own:
- distillation (cheap local character brains) + RL (self-improvement from in-game rewards) + multi-agent (the world) + eval (judge the agents).
- It’s “build everything Mac-buildable” + “what comes after autonomous agents (self-improving + multi-agent)” + “an environment most people have to build.”
What’s needed to make it runnable
The game’s repo specifics:
- How does a character pick an action each turn? (a model call? an OpenAI-compatible endpoint? what observation/action format?)
- What in-game signals can serve as rewards?
With those, scope the trajectory-instrumentation + the first GRPO experiment.
Related
- model-vs-agent — the static picture (what makes a model an agent).
- advanced-llm-training §13 — GRPO mechanics.
- agent-context-hierarchy — context discipline for the loop.
- Distillation result (cost-compression: 0.6B ≈ 4B on tool-calling) — the cheap local brains.