BuildSeniorData / MLAuto-graded

Build a ReAct loop that reasons then acts

An assistant must parse its own written reasoning, act on it, and recognize a final answer, or it loops forever.Strip away the frameworks and almost every tool-using agent is the same small loop: the model reasons, picks an action, you run that action as a tool, you feed the result back, and it reasons again until it declares itself done. That is the ReAct loop, and writing a correct one by hand is the most clarifying thing you can do to understand agents. The one in react_agent.py is broken. Make it parse actions, feed observations back, terminate properly, and survive malformed model output.

40 min

python

Start lab

You're done when

The loop parses the action name and its argument out of the model's output.
The observation from running a tool is fed back into the next step.
A final answer is recognised, extracted, and terminates the loop.
Malformed action output is recovered from rather than crashing the run.
The loop respects its maximum step count instead of running indefinitely.

What it measures

Agent engineeringImplementing the reason-act-observe cycle correctly, including how the loop decides it is finished.

Context engineeringThreading observations back into the model's context so each step builds on the last.

Robustness to model outputTreating a malformed action as expected — models produce them — and recovering instead of failing the run.

Your environment

Terminal + editorA full shell in the browser

Claude CodePre-authenticated, no key setup

Isolated microVMYours alone, wiped afterwards

Start this lab

40 min

Start labYou'll sign in to Anthropos to start. The lab opens in a private cloud workspace with Claude Code ready to use.

AI Labs is part of Anthropos Workforce. Not on your plan yet? Book a demo