From Loop to Harness
The loop is call, tool, result, repeat. The harness is the function that decides what of that result the model is allowed to read. This note points at the loop, the harness, OpenClaw's flush, the MCP servers, and the note on judgment. It does not add a new mechanism.
From Loop to Harness
A short linking note. It only connects what the longer posts already say.
The loop note is the whole agent in one while. Call the model. If it asks for a tool, run it. Push the result back. Repeat until the reply is a normal message. In that snippet, content is JSON.stringify(result).
The harness note starts at that push. The string the model reads is a decision: a prefix slice, whole records plus an omission count, a file path and a preview, or a shorter tool list. The measured case is in that note, not here. 11,189 tool results, 3,827 of them cut, and the largest prompt at 5.75% of a 400k window. The window was not the constraint. The cut was.
OpenClaw’s memory flush is one hook on the same boundary. A silent turn writes what should survive to memory/YYYY-MM-DD.md before compaction summarizes the older messages. The harness note’s offloading section is the general shape: durable text in a file, a handle in the window. The two storage shapes, a memory MCP server and those Markdown files, are compared in Two ways to give an agent memory.
The MCP note is how the tool list gets large. Five servers, each a handful of tools, and the default is to mount all of them. The harness note’s tool-loadout section is the countermeasure: filter by tag, cap the list, because every schema is text the model attends to. The MCP note’s update records that the 28 Jul 2026 spec made the protocol core stateless and made list results cacheable. The catalog can be cached. Which tools to attach this turn is still the harness.
AI won’t replace you is the judgment side of the same split. That note says the model can propose several architectures and a person decides which one fits. The harness note says code owns the budget, the record boundary, the omission count, the file written before a summary, the flush, and which tools are eligible. The model owns the judgments inside that. Keyword plus vector, the search both the RAG layer and OpenClaw’s index use, is Hybrid search keeps showing up. Retrieval decides what is worth addressing. The harness decides what is addressable.
ciao