Michael creates isolated Git worktrees off the main branch for every task so multiple coding agents can run concurrently. While this setup prevents repo conflicts, developers must remember that shared machine resources like database connections and dev server ports remain unisolated.
Enforcing a strict separation between an orchestration layer and a reusable service layer stops agents from writing redundant functions across files. This structured design pattern keeps LLM-generated code readable for human reviewers and easier for other agents to interpret.
Requiring coding agents to record screen video or capture UI screenshots proving a bug existed and was resolved builds trust before manual evaluation. For backend features without a user interface, generating before-and-after performance metrics serves as objective verification.
Chaining an automated code review loop forces agents to iteratively address feedback until achieving a perfect review score. This automated cycle prevents developers from wasting time on half-baked pull requests by guaranteeing baseline code standards prior to human testing.
Instead of having AI draft formal specifications, Michael speaks out loud with the agent for several minutes to surface nuances and alternative approaches. The conversation history acts directly as the project spec, while setup guides are saved to markdown only for future reference.
Michael notes that Astra reliably discovers subtle vulnerabilities and performance gains without being throttled by safety filters on security topics. However, Fable remains superior at writing clean, production-grade code that passes human review without unnecessary refactoring.
State-of-the-art flagship models often overcomplicate basic tasks and consume excessive tokens compared to efficient smaller models. Engineers should reserve top-tier models for orchestration and code review while delegating routine knowledge or computation tasks to cheaper alternatives.
Scheduling autonomous agent jobs to run recurring tasks can quietly create thousands of lingering Git worktrees and orphan dependencies. Developers operating local machine clusters or cloud agents require dedicated cleanup jobs to prevent severe memory exhaustion.
Advice from AI frontier labs often presumes endless compute and token subsidies, which is impractical for production software engineers operating within real budgets. Michael warns developers to critically evaluate lab recommendations and explore open-weight models to avoid vendor dependency.
