Egor Urvanov

Harness

Harness · agent harness

Everything around the model that turns it into a working agent: instructions, tools, permissions, memory and checks. The model is the same, the result depends on the harness.

CLAUDE.mdharnessLLMmodelrules · CLAUDE.mdcontextmemoryskillstoolspermissionshookschecksmodel + harness = agent● model, people, systems, data  ● agent and harnessEverything around the model that makes it a working agent

What a harness is made of · 15

  • Agent loop · The repeating rhythm of an agent’s work: think, call a tool, look at the result, think again.
  • Tool and tool calling · A tool is an action the agent can call: read a file, find a ticket in Jira, run a database query.
  • MCP · An open protocol through which an agent connects to outside services — Jira, Airtable, databases — via one interface.
  • Context · Everything the model sees at the moment: the conversation, project files and data from connected systems.
  • Agent memory · Notes an agent keeps between sessions: project rules, constraints it found, agreements.
  • CLAUDE.md · A file in the project root with standing instructions for Claude Code: project layout, build and test commands, conventions.
  • Project bible · A description of how the project is built and why: system boundaries, agreements, principles and values: what matters more when choosing and what is never allowed.
  • Skill · A folder with a `SKILL.md` file: instructions and helper files for one recurring task, such as a report or a migration.
  • Permissions · AI agent permissions define what the agent may do in your environment and connected systems: only read or write as well, and which commands to run.
  • Sandbox · A fenced place where an agent’s actions run: only allowed folders are visible and only allowed network hosts are reachable.
  • Hooks · Scripts the harness runs automatically at set moments of the agent’s work: before an action, after a file edit, on finish.
  • Static checks · Tests, linters, builds and other automatic verifications that confirm the agent’s result before a person looks at it.
  • Human in the loop · Human in the loop (HITL) means the agent does not finish on its own: at the important steps it waits for a person’s approval.
  • Orchestration · Managing several steps or several agents: who is responsible for what, in which order they work and how results are passed on, through files, a subagent result or cross-session messaging.
  • Logs and tracing · A record of each agent step: what it read, which tools it called and what came back.

Source: Anthropic: Effective harnesses for agents

Read more: AI agent = LLM + harness: LLM vs harness vs agent

Where it appears · 11

More 6

Related terms

Whole glossary

© 2026 Egor Urvanov