building-with-ai-agents

Use the Junior Intern Test: delegate tasks requiring some judgment but not your full expertise, describable in one or two sentences. Give well-defined tasks with clear verification criteria rather than open-ended problems like re-architecting a codebase.

Assign distinct tasks to separate agent instances simultaneously, as Cognition's engineers do with up to five Devins each. Use asynchronous handoffs through Slack, Linear, or GitHub issues, and review the resulting pull requests as you would for junior engineers.

Use a /review command so the primary model checks its own code by severity, then have a second model like Codex or Cursor independently review it. A /peer review command frames the first model as dev lead defending or fixing the flagged issues.

Yes, using workflows like Zevi Arnovitz's: explore the problem with a command that fetches Linear context, generate a markdown plan, execute it, and use multi-model peer review to catch errors. The 4x4 Debugging Framework fixes broken code without coding knowledge.

Most agent failures are context problems: underspecified instructions or missing information. Encode tribal knowledge into markdown files, code comments, and documentation, and avoid vague prompts that assume the AI understands implicit intent.

Avoid them for open-ended problems without clear verification loops, such as full codebase re-architecture. Also avoid vibe coding where you ship generated code without understanding the implementation, which creates unmaintainable systems.