{"t0":"Four pairs. Four moves. Keeping this pebble.","t1":"A full set. A very good afternoon.","t2":"Give the AI a goal. The model can request edits and checks. The surrounding program, called a harness, runs those tools and returns their results.","t3":"During development, the AI author used a file-writing tool to add the matching rules. These buttons now use that game code. Try a pair.","t4":"The rules pass these checks. Watch a miss, one moment at a time. The pictures below come from the same code as the game.","t5":"That was the next task given to the AI author: keep the fewest moves from a finished round. The storage checks passed. Then review caught something else.","t6":"The score upgrade reset an active round.","t7":"Adding saved scores reset a round already in progress. We reproduced that bug in a browser and sent the result back to the AI author. Your game stays untouched here.","t8":"Repair the upgrade without losing progress.","t9":"The author added a test that caught the reset, then changed the code. Turn over a shell above. Apply the corrected edit and see whether it stays put.","t10":"Feedback changed the next edit.","t11":"The upgrade now keeps the current round. This is the loop: an action returns a result, the result reaches the model, and the model can choose another action.","t12":"You give the AI a task. It can ask tools to edit or check something. The surrounding software runs allowed requests and returns what happened. That software is the harness.","t13":"During the recorded build, the AI used a tool to write the matching rules. The game above now runs that code. Try turning over two shells.","t14":"These checks pass for the game rules they examine. Watch what happens after a wrong pair. The pictures use the same code as your game.","t15":"The next recorded request was to save the fewest moves from a completed round. Tests of the saved score passed, but review found a different problem.","t16":"The new score feature accidentally restarted an unfinished round. We reproduced the bug and sent the result back to the AI author. This replay does not reset your game.","t17":"The author added a test that exposed the reset, then fixed the code. Turn over a shell and apply the corrected edit to check whether the shell stays open.","t18":"The corrected upgrade keeps the active round. The model received the earlier failure and used that feedback for its next edit. This is a replay of that work, not a live model.","t19":"This is the recorded next task, not a new live request.","t20":"Condensed build replay. The game and checks run here; the AI is not live.","t21":"Your round survived the upgrade","t22":"Compared the board, revealed cards, moves and found pairs immediately before and after this upgrade.","t23":" The round check used your board at upgrade time.","t24":"A check failed. Keep the result.","t25":"This replay cannot make a new repair. A live agent would need to read the failure before choosing its next attempt.","t26":"This replay cannot invent a repair. A live agent would receive this result before deciding what to try next.","t27":"Some things are worth remembering.","t28":"Find four pairs. Keep your best.","t29":"Turn over two shells and look for a pair. Finish a round to save your best move count, if this browser allows storage.","t30":"Turn over two shells at a time. Your fewest moves from a completed round are saved when browser storage is available.","t31":"One otter. One good pebble. A game to make.","t32":"This one has a nice weight to it.","t33":"A tool doesn’t have to be complicated.","t34":"Still the same pebble. Still a good choice.","t35":"Game file created. Open otter-shell-pairs.html from your downloads to play. No network connection is needed.","t36":"The download did not complete. Your game is unchanged; try again.","t37":"Run a check for these settings. An earlier result does not verify them.","t38":"The check did not run. No correction is verified.","t39":" A fresh round cannot reveal lost work. Try a shell already open.","t40":" Select Keep the round and rerun the same case.","t41":" Restore the result and rerun.","t42":" Continue to apply the existing correction to your game."}

Inside an AI agent / 01

How an AI agent
checks its work.

Play the shell game, follow its recorded AI-assisted build, and try the checks that caught lost progress.

Recorded build. Playable result.

A small corner of the ocean

One otter. One good pebble. A game to make.

The thing we’re making

Shell pairs

Preview

0 of 4 pairs

0 moves

Use Tab or the arrow keys to move between shells. Enter or Space turns one over.

Keep missed pairs open until I’m ready

The shells are here. The game isn’t wired up yet.

The build, replayed

Recorded steps. No live AI.

1 Model requests
2 Tool acts
3 Result returns to the model
Feedback becomes the next input

Start with a request

“Make a shell-matching game.”

You give the model a task. It can ask tools to edit or check something. The software around it runs allowed requests and sends back what happened. That surrounding software is called a harness.

A condensed account of this game’s development, with checks you can run.

The useful bit

How tool results guide the next edit

The helper needs to see what happened before deciding its next step. A failed check can lead to another edit. A limit on turns stops it from trying forever.

Download the shell game

Save one HTML file to play offline and edit.

The finished game, not an AI agent. No account or API key.

What is real here?

You can play the game and run its checks here. The build story retells an earlier session made with AI assistance. It is not a live AI conversation and does not write new code as you watch.

The saved-score change began as a second task. Review then caught a real bug: the upgrade reset an active round. A browser reproduction was returned to the author, who added a failing regression test and corrected the code. The replay summarizes that exchange; it does not show private model reasoning.

The harness is the software around a model: it supplies tools, runs allowed requests and sends results back. These teaching panels describe that relationship; they do not run a model harness.

Created by Jonathan Simone with AI assistance. Original artwork and game. Technical background: Building effective agents and the action and observation cycle.

How an AI agentchecks its work.

Shell pairs

“Make a shell-matching game.”

A miss should close.

How tool results guide the next edit

Download the shell game

How an AI agent
checks its work.