A Python framework for orchestrating two to ten language-model agents and proving what the collaboration buys. Each agent's knowledge lives in a 256-dimension cognitive model, a learned gap detector decides when a difference is worth a conversation, a PPO policy decides who speaks, and convergence is detected by entropy.
Four learned components sit under the orchestration. A 512-dimension recurrent state encoder embeds the dialogue with the task semantics and collaboration history. A 256-dimension bidirectional knowledge model with attention tracks what each agent knows round by round. A four-layer gap detector, narrowing 512 to 64, decides whether two agents should talk or keep working alone. A PPO actor-critic policy chooses which agents to select and how to split the task. Convergence is entropy-based, aggregation is confidence-weighted, and deadlock detection breaks communication loops that will not close.
Three primitives prove whether collaboration was worth its cost. An evidence lab runs learned, random, independent, debate and no-model strategies over the same tasks with the same seeds and emits raw JSONL plus cost, quality and latency summaries. A cognitive trace writes append-only, hash-chained messages and decisions that a command verifies after the fact. An adaptive budget policy chooses between adding an agent, critiquing, synthesising and stopping, inside explicit token and round limits.
Embedding calls are batched per round: messages deduplicated within the round, cache entries namespaced by model and input type, at most one provider request per 96 unique texts, and vectors restored to the exact order the inputs arrived in — with a test covering the ordering and the failure semantics.
The benchmark suite is wired for MultiWOZ, CoQA, HotpotQA, CommonsenseQA and MMLU at three, five and seven agents, and training runs supervised pre-training with paraphrase and back-translation augmentation before reinforcement learning.