Why would you want to do this?

• Longer independent workflows: hand over a task, not a question
• Feedback comes from the environment, not from you
  • Coding: build, run the tests, read the errors, fix – the test suite is the harness
  • Science: run the analysis, inspect the output (plots, summary statistics, sanity checks), revise
• You review results and key decisions, not every step

The longer the leash the higher the risks

• Unintended actions
  • deleted or overwritten files and data
  • leaked credentials or unpublished data
  • rogue actions on shared resources: repos, clusters, email
• Wrong results
  • plausible output that is wrong
  • insufficient tests, or the agent adapts the tests until they pass
  • more output than you can verify
• Code as liability
  • large, complex code base that nobody understands well enough to maintain
  • bugs surface later, with no one who knows where to look

Mitigation: contain, verify, simplify

• Contain
  • sandbox: agent writes only to the project directory
  • no credentials or sensitive data within reach, restricted network
  • everything recoverable: version control and backups
• Verify
  • checks you define: known results, positive controls, tests you have read
  • inspect intermediate outputs, not just the summary
  • scope tasks to what you can verify
• Simplify
  • ask for less code, reuse established libraries
  • review diffs, insist on simple architecture
  • treat it as code you are responsible for

Example: fitness effects of Spike mutations on structure