Repair a bug and check the cases a single example misses.
Change the briefRun working toolsInspect the result
YOUR TURN TO EXPERIMENT
Free to explore · No sign-in or API key
Give the agent a brief.
Investigate a broken ticket calculator, choose a repair strategy, and execute tests before and after the patch. Review the changed code alongside its results.
Make the calculation work for different orders.
Ready when you are
The tools calculate and check your inputs in an isolated JavaScript runner.
Working local tools. Authored controller and classroom data. No live AI model or external action.
A SMALL BRIEF. A VISIBLE RESULT.
See the agent at work.
Start with a scenario, change a detail, and run the demo. Your result and the steps behind it will appear here.
1 · Read the brief2 · Use tools3 · Check the result
Try this twistFirst, change the order and rerun the calculation. Then try the hard-coded patch and identify which tests reveal its limitation.
Open the code behind this agentRead it, change it, run your own variation.
This is the JavaScript that powers the demo. agent.js coordinates the workflow; tools.js contains its available tools. Tool calls and returned data are inspectable after each run.
Ready when you are
⌘ / Ctrl + Enter runs your agent.js draft with the current inputs. Tab leaves the editor. Package imports and network calls are unavailable in this bounded environment.
The real research behind this demoWhere the idea comes from, and what this classroom version does.
Anthropic's coding-agent guidance connects issue investigation, code changes, test feedback, and human review. This working repair exercise lets you inspect each part of that software workflow.
The JavaScript and tests really execute in the sandbox. Patch strategies are authored examples; this is not a connection to Claude or a benchmark result.
STEMtiq authored this exercise. The named organizations do not endorse it. The JavaScript tools run locally; the controller is programmed, and classroom records are labeled examples. This is not a live instance or reproduction of the referenced research system.
Anthropic
Building effective agents
Anthropic describes tool-observation loops, bounded execution, and coding agents that use executable tests as feedback.
Our deterministic controller runs a tiny repair workflow. It is not Claude, an LLM-directed agent, or a reproduction of benchmark results.
Your JavaScript really runs. The model decisions and school data are authored simulations, so you can learn without an API key. Every workspace also includes a separate real SDK example to explore next. Passing the lab’s cases is practice, not proof that an agent is ready for real-world use.