Skip to content

What Is a Harness

In current usage, an agent harness is everything wrapped around the model except the model itself: tool execution, memory and state, orchestration, guardrails, and observability.

Agent = Model + Harness

A good harness is what turns a general LLM into a system that can run long, complex workflows reliably. It is the difference between a model that quits after one turn and one that finishes the job.

The six layers

flowchart TB
    API["API Layer<br/><small>entrypoint, streaming</small>"]
    ORC["Orchestration Layer<br/><small>plans, loops, replanning</small>"]
    TOOL["Tooling / MCP Layer<br/><small>tools, skills, schema</small>"]
    MEM["Memory / Context Layer<br/><small>logs, durable artifacts, RAG</small>"]
    EXE["Execution / Sandbox Layer<br/><small>run code, isolation, quotas</small>"]
    OBS["Observability Layer<br/><small>events, traces, audit</small>"]
    API --> ORC --> TOOL --> EXE
    ORC --> MEM
    ORC --> OBS

Where Leopold sits

Claude Code is already a strong harness for a single interactive turn. It has the tooling, execution, and a permission system. What it lacks for unattended, long-running work is exactly what Leopold adds:

flowchart LR
    subgraph CC["Claude Code (one turn)"]
        T[Tooling] --- E[Execution] --- P[Permissions]
    end
    subgraph LEO["Leopold (long-running)"]
        DEC["Decider<br/>(your charter)"]
        CON["Continuity<br/>(stop hook / driver)"]
        GUARD["Guardrails<br/>(git lock)"]
    end
    CC --> LEO
  • a decider (your charter) so the agent has authority to choose instead of asking,
  • continuity (a stop hook, or the SDK driver) so a finished turn rolls into the next item,
  • guardrails (a tool-call gate) so autonomy never touches your git.

See the full mapping in the Architecture Overview.