A model turns text into text, and nothing else. It is the sealed engine in the middle: text goes in through one slot, text comes out of
another, and it remembers nothing between calls. Every call re-reads the whole tray in front of it, the context window,
from the start. Everything around it is the harness: if you’re not the model, you’re the harness.
The harness does four jobs.
Every part, pulled out one at a time
The same task with one part removed. Removing a part also stops the parts that stand on it. Click a row to load that factory.
The parts that save the most tokens (subagents, offloading, skills) keep large text out of the window. The parts that cost few or
no tokens (the sandbox, the gate, planning) stop a run from doing the wrong thing. A harness is judged on both.
What the figure leaves out
- The numbers are illustrative. Token sizes match common orders of magnitude for a coding agent; they are not measurements of any product. The run is scripted, not a live model. The arithmetic over them is exact.
- One path through the task. A real agent branches and sometimes recovers on its own. Here each missing part has one fixed consequence.
- Prompt caching is not modelled. A cached prefix lowers the cost of carrying stable instructions; it does not shrink the window.
- More parts exist. System prompts, tool definitions, MCP servers, hooks and middleware, model routing. The four jobs are the frame, not a complete parts list.
- Steering is more than a gate. A person can also interrupt and redirect a run mid-way.
Sources
- Vivek Trivedy, LangChain, The Anatomy of an Agent Harness, 10 March 2026.
- LangChain docs, Agents: execution environment, context management, subagents and steering.
- Anthropic, Effective context engineering for AI agents, 29 September 2025: compaction, note-taking, sub-agents, progressive disclosure.
- Manus, Context Engineering for AI Agents: Lessons from Building Manus: the file system as restorable context.
Part of the Agents & Workflows category on aiengineer.help.