Agent Loop
Streams model output, detects tool calls, executes work, observes the result, and feeds evidence back into the next turn.
Open agent infrastructure
Agent Harness
Agent Harness is the operating layer around an LLM: tools, skills, memory, permissions, observation, and coordination that turn model output into reliable, reviewable work.
- Tools
- Files, shell, web, MCP
- Memory
- Context and session state
- Safety
- Permissions and hooks
Definition
What is an Agent Harness?
An Agent Harness is the complete infrastructure that wraps around a language model and lets it behave like a working agent. The model provides intelligence; the harness provides controlled action, observation, memory, and operating boundaries.
OpenHarness frames this idea as lightweight core infrastructure for tool-use, skills, memory, and multi-agent coordination. This site summarizes that pattern for builders evaluating how agent systems should be assembled.
In practice, the agent harness is the difference between a model that can answer and a system that can inspect files, call tools, enforce approvals, preserve context, and recover from failed steps.
Core components
The harness is the part that makes agency operational.
Harness Toolkit
Exposes file, shell, search, web, MCP, and domain-specific tools through predictable interfaces the agent can call.
Skills and Plugins
Loads specialized instructions, hooks, prompts, and extensions only when they matter for the current task.
Context and Memory
Maintains project instructions, session history, persistent notes, compaction, and resumable task state.
Governance
Applies permission modes, path rules, command rules, approvals, and pre/post tool hooks before side effects happen.
Coordination
Delegates tasks, tracks background workers, manages team state, and merges results into the primary agent session.
Architecture flow
A useful harness closes the loop between intention and verified action.
- PromptIntent enters the session
- ModelReasoning selects the next action
- PolicyPermissions check the proposed tool call
- ToolWork runs in a controlled surface
- ObserveOutput returns as evidence
- MemoryState persists for the next step
Reference implementation
OpenHarness shows the pattern in code.
HKUDS/OpenHarness is an MIT-licensed Python implementation of agent harness infrastructure. It supports provider setup, a terminal UI, tool execution, skills, memory, permissions, MCP integration, and multi-agent coordination primitives.
FAQ
Agent Harness basics
Is an agent harness the same as an LLM?
No. The LLM is the reasoning engine. The harness is the surrounding runtime that gives it tools, state, rules, and feedback.
What should builders evaluate first?
Start with tool boundaries, permission behavior, observability, memory strategy, and whether the loop can recover from failed actions.
Which parts matter most for a production agent harness?
Tool contracts, permission policy, memory retention, logs, retries, and the release process around the harness all matter before adding more model complexity.
Is this website affiliated with OpenHarness?
No. It is an independent Agent Harness keyword reference that links to OpenHarness as a public source.