← TinyGPT · docs · devlog · roadmap · speedup
source: docs/learn/rl-multi-agent-roadmap.md · view on GitHub ↗

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 ingredientIn the game
Policythe character’s model brain — the distilled small local model (cost-compression lever)
Environmentthe world
Rewarda 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)

  1. Pick one measurable behavior (a character completing a specific goal / winning a negotiation).
  2. Instrument the game to log each turn (observation → action → outcome→reward).
  3. Run GRPO on a small local model (the distilled specialist) as that character’s brain; reward = the in-game outcome.
  4. 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)

Why this fits the north-star

It fuses everything into one Mac-runnable system, on an asset we already own:

What’s needed to make it runnable

The game’s repo specifics:

With those, scope the trajectory-instrumentation + the first GRPO experiment.