New forms of intelligence are evolving at an accelerating rate, yet the most widely used models remain fundamentally constrained—locked into human-defined training cycles, reliant on static pattern recognition, and often incapable of true independent thought. To unlock intelligence that transcends pattern matching, AI must move toward end-to-end autonomy, where these entities can not only execute tasks but also train, refine, and expand their reasoning capabilities without human intervention.

Among the emerging solutions, Daydreams ($Dreams) stands out as a generative agent library designed to push the field beyond rigid learning pipelines. Unlike many frameworks that depend on supervised fine-tuning (SFT) models, @lordOfAFew’s creation integrates reinforcement learning (RL) to enable continuous self-improvement, strategic decision-making, and long-term hierarchical task execution. By developing AI that can think, plan, and evolve autonomously, Daydreams lays the foundation for systems capable of original thought and human-level creativity—exactly what’s needed to tackle the seemingly unsolvable problems facing humanity today.

The Generative Approach to Autonomous Agents

Daydreams is an open-source, MIT-licensed generative agent framework written in TypeScript. It’s designed to support cross-ecosystem operations on Ethereum, Solana, Base, Arbitrum, Hyperliquid, Optimism, StarkNet, Abstract, Mud.dev, and Dojo. Unlike many AI agent frameworks, which rely on predefined workflows, this one follows a generative approach. Instead of executing pre-scripted interactions, its agents intelligently generate and carry out actions based on contextual information.

This flexibility is especially valuable in on-chain environments, where conditions change rapidly and require agents that can pivot in real time without human intervention. By enabling agents to ad