Coding, Computer-Use & Deep Agents

Run ReAct agents in web, code and computer-use environments and see where the loop changes.

Lessons

5 lessons

Deliverable: Coding agent: give a deep agent a developer sub-agent and a tester sub-agent, ask it for a small Python function with tests, and log each iteration until the tests pass.
Assessment: Answer the five Interview questions without notes.
Interview probe: You ship a deep agent to Larkspur's support engineers, built like `03b_long_term_memory_agent.py`. Someone sets `USE_SANDBOX=true` expecting cloud isolation, and the memory panel shows one engineer's notes in another's session. What is wrong?
Common mistakes: Choosing a ReAct agent for a fixed-shape weekly report instead of a planned pipeline
Running a vision-driven computer-use agent with no step cap
Assuming USE_SANDBOX=true runs commands in a sandbox rather than LocalShellBackend
Trusting a tool-less tester agent's reported pass rate as executed test results

How confident are you with this module?

1 means "I could not explain it yet"; 5 means "I could teach it". Rate again any time.

Your notes on this module

Saved in this browser when you leave the box.