Harnesses

The box around the loop

Lesson 0006 gave you an agent: a while-loop where the model picks the next tool. A harness is not a second model. It is the software you wrap around that loop so the program is allowed to run in the real world.

1. 1 · AllowMay this slip run? Sandbox, or a human stamp.
2. 2 · RunExecute. On failure, paste the error. Retry.
3. 3 · StopDone, MAX steps, timeout, or pause for a person.
4. 4 · ShowThe plan is visible. Do not hide the loop.

Allow · run · stop · show. The inner loop from 0006 does not change.

The loop is small. The product is the box.

Anthropic: an agent is typically just an LLM using tools in a loop — and you still need stopping conditions, sandboxed testing, and a pause for a human at checkpoints. 12-factor agents says the important gap is between the model writing a slip and your code running it. Building effective agents · Factor 8 · Own your control flow

Same loop as lesson 0006. Architecture tonight = what wraps it.

# NAKED LOOP — lesson 0006, nothing around it
while steps < MAX:
    action = llm(messages, tools)
    if action is text:
        return action
    result = run(action)          # slip → till
    messages.append(result)

# SAME LOOP, INSIDE A HARNESS
while steps < MAX:                # STOP is your rule
    action = llm(messages, tools)
    show(action)                  # SHOW the plan
    if action is text:
        return action
    if not allow(action):         # ALLOW: sandbox / stamp
        messages.append("denied")
        continue
    result = run_with_retry(action)  # RUN: errors become paste
    messages.append(result)

Kitchen only to make the gap concrete. The load-bearing picture is the box and the two programs.

Skill: name which job fired

Same agent loop. Four wrappers. On Allow, the loop waits for your stamp — that click is the harness, not the model.

You have been sitting in one

ChatGPT, Claude, Cursor, Claude Code: you talk to an agent loop someone else already wrapped. File permissions, the stop button, “run this terminal command?”, retries when a tool fails — that wrapper is the harness. Lesson 0006 asked whether Cursor is an agent or a harness. Both: an agent inside a harness. At work you will write (or pick) the wrapper, not just the while.

Using a product is not the same as shipping the box.

Buy a harness, or write the four lines

LangChain sells the split as layers: a runtime (durable execution, pause/resume), a thin loop, and an opinionated harness on top. Anthropic’s advice still stands: start with the API; if you use a framework, know what it is wrapping. Deep Agents’ extra furniture — filesystem, skills, sub-agents, memory — is a fat harness. We will unpack those pieces later. Same LangChain essay · Anthropic, “When and how to use frameworks.”

Anthropic’s three agent principles, now that you have a harness to hang them on: keep the design simple; show the plan (Show); invest in the tool interface the model sees (Allow + the slips from lesson 0005). Compaction and skills are how a harness spends the window — next lessons, not tonight.

Check