The Memory is the Harness: Why Your Moat Isn't the Model

The Anatomy of a Harness: A deep dive into the 512k+ lines of code that make an agent functional—filesystems, bash tools, sandboxes, and orchestration logic. 

Memory as Context Management: Why the harness is responsible for the entire lifecycle of context, from loading AGENTS.md to managing server-side compaction and session persistence. 

The Lock-in Trap: How closed harnesses create proprietary data flywheels for providers, making it impossible for you to swap models without ""lobotomizing"" your agent's learned user preferences. 

Harness Engineering vs. Prompt Engineering: Shifting from ""asking"" the model to be better to ""engineering"" the environment so the system physically cannot repeat a failure. 

The Path to Open Memory: Strategies for building open, model-agnostic harnesses that ensure you own your proprietary dataset and your agentic ""moat."" "

"The industry is reaching a consensus: Agent = Model + Harness. While the LLM provides the reasoning, the harness provides everything else—the state, the tools, and the memory. But as model providers move toward closed, stateful APIs, a critical risk has emerged. If your agent’s memory is locked inside a proprietary harness, you aren't building a product; you are renting a relationship. 

This talk explores why Harness Engineering is the most critical discipline for AI practitioners in 2026. We will break down the anatomy of a harness and explain why memory isn't just a plugin—it is the harness.