Plan mode is dead
As AI coding assistants like Claude and Codex improve, many users argue that a separate “plan mode” — where the model outlines steps before editing code — has become redundant, since newer models can be instructed conversationally to plan first and then act. Others strongly defend explicit planning phases as essential guardrails: they surface design assumptions, keep humans aligned with what the agent will change, provide an audit trail, and reduce the risk of costly or dangerous edits (for example to schemas or production databases). Across both views, there is broad agreement that planning itself remains crucial for complex or high‑blast‑radius work; the real debate is whether it should live in a dedicated mode or as part of more flexible workflows using prompts, spec files, and multi‑agent review.
Role and Value of Plan Mode
- Many developers still rely on plan mode as a distinct “pause before acting” step:
- Forces the model to gather context and outline changes before touching code.
- Lets humans preview architectural decisions, risks, and tradeoffs before expensive or risky edits.
- Serves as a dry run or “audit checkpoint” for high‑blast‑radius work (auth, payments, schemas, prod DBs, complex refactors).
- Some use plan mode to iteratively harden specs: rerun until questions stop, then execute.
- Others treat plans as durable artifacts (plan.md, spec directories, AGENTS.md) that can be versioned, reviewed by multiple agents, and reused across sessions/models.
Arguments That Plan Mode Is Obsolete or Overrated
- Several users say newer models (e.g., recent Claude/Fable/Opus, some GPT/Gemini setups) “just get it”:
- A simple instruction like “plan first, don’t write code yet” in normal mode works reliably for them.
- Long, generated planning prose is seen as hard to review and misaligned with how humans actually build understanding.
- Some claim the original need—preventing “loose cannon” agents—has diminished as instruction‑following improved.
- There’s concern that built‑in plan modes are just harness prompts, not true sandboxes or permission systems, and may add cost and latency without proportional benefit.
Alternative Workflows and Tools
- Common alternatives:
- Conversational planning with explicit “no edits yet.”
- Dedicated plan documents maintained by agents, with fresh sessions for implementation.
- Multi‑agent setups: one model plans, others implement, others review (sometimes via separate tools like plannotators, “grill me” skills, or MCP servers).
- Spec‑driven or blueprint workflows where humans write or tightly curate specs and agents fill gaps.
Control, Safety, and Quality Concerns
- Strong anxiety about unintended edits, especially to databases and large codebases; some want explicit read‑only or sandbox modes.
- Fear that “vibe coding” without structured planning accelerates tech debt and erodes developers’ understanding of their own systems.
- Counterpoint: even if understanding lags, agents can later explain, diagram, and refactor AI‑written code, shifting emphasis from manual ownership to verification and oversight.
Overall Sentiment
- Broad agreement that planning is essential.
- Divided views on whether a dedicated “plan mode” UI/mechanism remains the best way to achieve it.