Agent harness
The structure around a model that turns it into one working agent: its interface, agent loop, own context and memory, permissions, and list of allowed tools.
The test for what belongs in it: does each agent carry its own, or do they all share one? If each agent has its own, it's harness. If they share it, it's stack. A harness wraps an agent, not a person. One person can run several agents, each with its own harness, and an agent can run with no person attached at all. Your personal assistant is one agent with one harness; that it's yours is about who it acts for, not about the harness belonging to a person. Why the agent is the unit: if the line were drawn around a person, a coding agent and a research agent run by the same person would be lumped into one harness, even though their loops, tools and permissions are completely different. The test stops working. Components that sit on the line: - Tools: which tools an agent may use is harness; the shared connector behind each tool is stack. A tool local to one person's setup is harness; a centrally hosted one is stack. - Permissions: the policy for one agent is harness; the central identity system that enforces it is stack. - Gateway: stack, because it is shared. People disagree on this one, so say it explicitly. Harness vs stack: the stack is the shared foundation: gateway, inference, data and ontology, evals, central identity, and shared connectors. The same harness concept appears in both the build stack (e.g. Claude Code: terminal, files, shell, git, human in the loop) and the product runtime stack (e.g. a support bot: narrow interface, business systems, tight permissions, little oversight). Same components, different tuning. Promotion is how something moves from a harness into the stack. The word grew up around personal coding agents (Claude Code, Codex, OpenCode), but it applies to customer-facing AI too: a product owner picks a harness for a chatbot. Choosing Claude Code vs OpenCode is mostly a harness decision, not an interface or model decision, since they can call the same models.
Also called
- Scaffold
- Scaffolding
- Agent runtime
- Harness
- Agent infrastructure
Avoid
Harness engineering; Agent framework
Kinds
- System prompt A prompt written by the developer or vendor, not the user, that the app sends alongside every user-supplied prompt to set the model's role, rules, and tone.
- Agent skill A package of instructions and files, written in an open text-based format, that an agent loads only when a task needs it.
- Agent loop The cycle an agent repeats: decide the next step, take it with a tool, look at the result, and go again until the job is done or something stops it.
- Client The software that opens the connection and makes the call, to a model or to an MCP server.
- Interface The way a person interacts with an AI system: free text, document and image upload, audio, buttons and forms, in a chat window, terminal or IDE.
Related
- Orchestration The coordination and triggering of agents or steps: deciding what runs when, passing results between them, and handling failures.
- Loop engineering A practice of building where an agent keeps rerunning and checking its own work, with set rules for when it stops and when it asks a person to step in.
How it connects
- Agent harness turns a model into Agent
- Agent skill loaded into Agent harness
- Agent harness runs on Stack