Built-in web search in the OpenAI and Anthropic APIs is not a model capability. It is a small agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → sitting behind the API that calls a search tool for you — the same architecture you would write yourself, just hidden.
Storing conversation state on a provider's server (stateful APIs, server-side context compactionSummarizing an agent's earlier conversation to free room in the context window so a long session can keep going.Full definition →) quietly welds you to that model. You cannot switch providers and resume an old thread, because the thread lives in their format.
Scaffolding did not shrink as models improved, it moved. Simple retrieval chains gave way to graph-shaped flows, which gave way to harnesses. An AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → is an LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → touching tools and data, so something always has to sit between them.
Standalone memory services do not fit yet because the industry has no shared abstractions for memory. Long-term memory rarely makes the first version of a product: teams get the agent working, then worry about personalization.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
Why it matters
Why the harness — and specifically who controls your agent's memory — is critical infrastructure, and why open harnesses matter. A short, sharp argument for not outsourcing the most important part of your agent.
Key quotes
“When Claude Code’s source code was leaked, there was 512k lines of code. That code is the harness.”
“Asking to plug memory into an agent harness is like asking to plug driving into a car.”
Sarah Wooders
“If you use a closed harness, especially if its behind an API, you don’t own your memory.”
