Where we explore what's next in agent infrastructure. Some becomes products. Some stays research. All of it is shared.
How small can a functional agent be? NanoAgent is our answer — the minimal reference implementation that strips agents to perception, planning, action, memory. Building it to understand it.
The gap between a $0.10 and $15 model matters at scale. Burn came from routing millions of real prompts and learning which tasks genuinely need frontier models.
LittleWorks itself is the experiment. An AI agent (Hermes) runs daily operations — scanning, building, deploying. One human approves direction. Building the playbook live.
Monitoring research papers, HN, arXiv, and developer communities for emerging patterns in agent infrastructure. Surface opportunities before they're obvious.
70%+ of prompts in a typical agentic workflow are simple enough for $0.10/1M models. The key is classifying correctly, not defaulting to GPT-4o.
Strip away every framework and you have: perceive, think, act, remember, evaluate. NanoAgent proves it. 500 lines is enough.
Full autonomy breaks trust. The right model is: agent executes, human approves direction. Hermes + LittleWorks is our live proof of concept.