What is an AI harness?

Put a photo in
the right album.

The picnic photo should go into Family weekend. Let the helper look up the right album, or disconnect that tool. Watch whether the photo actually arrives and whether a receipt confirms it.

Scripted helper. Local tools. No AI calls or uploads.

Your shared albumsSample 07

Beach picnicBelongs in Family weekend

Held here

Family weekend

No copy added

Work event

No copy added

Your photo is ready. Both albums are empty.

Original illustrated sample. These albums exist only on this page.

TaskAdd the photo to its assigned album.
Requested actionNo request yet.
Returned observationNothing returned yet.
Visible resultNo copy added.

You control what becomes possible. The program around an AI model supplies its tools, observations and limits. That program is its runtime harness. Inspect the controls to see each job.

Inspect the controls

Inspect the harness and change its controls

The tools behind the album

Cutaway / inspect every part

This helper follows a written rule: read the assigned album, then add a copy. Both tools work only inside this page. No photo is uploaded.

Scroll the cutaway sideways, or inspect a part below.

Task & context: the harness supplies the task, tool list and request budget.

With lookup connected and a budget of two requests, lookup returns A and placement returns a receipt. Unplug lookup: the request is blocked before execution. Allow one request: lookup succeeds but placement cannot run. The decisions alone cannot add the photo.

01 / Follow the consequence

Compare the two runs

The reference run completes two requests. Change one control to compare it with your run.

Reference Connected / A / 2 requests

Request lookup → execute → observe A.
Request placement → execute → observe receipt.
Validate receipt → stop.

Your machine Same setup

Try unplugging the tool to see where the trace stops.

Try a worked example

Unplug lookup. The request still appears, but execution does not. Restore the socket, then lower the budget to 1. Now lookup runs and returns A, but placement cannot run. The harness controls which requests become actions.

02 / The receipt is not the verdict

It finished. Did it work?

The runtime checks that a receipt matches a request. An evaluation harness asks a different question: did this photo reach its assigned album?

Try “Wrong label: B instead of A” above. The machine can finish every step and still put the photo in the wrong album.

This evaluator checks a known toy world, not a real photo service or general AI quality. A sandbox would restrict where code can act, such as limiting files and network access. The drawn housing and permission switch are not a security sandbox.

03 / Try it somewhere else

The archive assistant lost its receipt.

A report belongs in cabinet A. The filing tool ran, but its return never arrived. Should the harness announce success, retry immediately, or stop with the result unconfirmed?

Reveal the explanation

Stop unconfirmed, then reconcile the actual state. A request is not proof of execution, and execution without a return does not supply a validated receipt. In real tools, a blind retry can repeat an action. This toy does not implement recovery or idempotency.

04 / Keep the machinery

A harness small enough to read.

Keep a browser workshop with your chosen setup. Open it with a double-click, change the controls and run it offline. No installation needed. Both tools execute in memory. The fixed rule can react to A or B; the fixed transcript stops when its expected observation changes.

The browser workshop is a self-contained HTML file. The optional code-only demo needs Node.js, with no packages or key. Run node harness-demo.mjs. Edit CONFIG inside the file to try another setup.

Original teaching code: MIT license

Sources, scope & credits

Jonathan Simone, independent developer at Simone Systems Research, with AI assistance. Original synthetic task and artwork. No model outputs or private project data are reproduced.

Anthropic: separating harness, session and sandbox informs the distinct controls.
LangChain: agents, tools and middleware informs the surrounding runtime jobs.
LangSmith: evaluation concepts informs comparing a run with an expected result.
Bret Victor: Up and Down the Ladder of Abstraction inspired keeping individual events beside complete traces.
Nicky Case: LOOPY inspired direct intervention and a useful export. No code or assets copied.

Sources inspected October 11, 2026. This is a deterministic teaching simulation, not a benchmark or evidence that harness changes generally improve model performance.