Embracing the parallel coding agent lifestyle (simonwillison.net)

An engineer reports switching to a “parallel coding agent” workflow—running multiple instances of Claude Code, Codex CLI, and cloud agents concurrently across different checkouts—to speed up non-blocking development tasks. They found this isn’t about having more code land faster (review remains the bottleneck) but about churning many low-risk or investigatory jobs in the background: proofs-of-concept, quick codebase explanations, grep-driven code navigation, and small maintenance fixes like silencing test warnings. Modern agents can even checkout repos, read new libraries’ source, run tests, and produce actionable patches, making them useful research and scaffolding tools rather than full autonomous committers. Key technical patterns and tradeoffs: use parallel agents for isolated, low-cognitive-overhead tasks; prefer carefully specified prompts for reviewable work; stash LLM-generated explanations as context for future prompts. Practical safety measures include running agents in separate directories or /tmp checkouts (worktrees not yet used), isolating risky runs in Docker, and using asynchronous/cloud agents when you permit network access (with known leakage risk). The author cites Claude 4 and GPT-5 as pivotal model improvements and recommends experimenting to build instincts for which tasks are safe to parallelize, while iterating on isolation and review practices as agent capabilities evolve.