The Problem: Velocity Sickness

Engineers using AI often experience 'velocity sickness'—a state where individual output increases rapidly, but team cohesion and product impact suffer. This manifests in four specific ways:

  • PR Overload: The team generates more code than can be effectively reviewed or merged.
  • Directional Fragmentation: Engineers sprint in conflicting directions because agents allow them to work in isolation without shared context.
  • Agent Bankruptcy: Engineers lose hours of work by repeatedly discarding agent sessions, leading to redundant token usage and wasted effort.
  • Loss of Agency: Critical product decisions are offloaded to agents, causing engineers to lose ownership of their own software.

Shifting from Implementation to Decision-Making

Traditional IDEs were built for the 'implementation and polish' phase of software development. However, AI has shifted the core work to a 'decision layer'—a creative, exploratory phase that requires different tooling.

To regain control, teams must move away from ephemeral, chat-based workflows toward doc-based workflows. Chats are isolated and 'brain-off,' whereas documents act as a durable, shared portal to the software system. By centering work around a document, engineers can:

  • Extract State: Keep agent context stateless by using the document as the source of truth, allowing multiple agents to collaborate on the same project state.
  • Prioritize Ideas: Use the planning phase to explore ideas before committing to implementation. This shifts the metric from 'code velocity' to 'idea velocity,' ensuring only high-impact features are built.
  • Enable Collaboration: Treat the plan as a multiplayer artifact. Sharing plans with teammates early allows for human-in-the-loop feedback before an agent spends hours building the wrong thing.

Immediate Action Plan

To implement these changes today, adopt these three practices:

  1. Separate Gears: Explicitly distinguish between the 'planning' phase and the 'polish' phase. Recognize when you are drifting between them and ensure your tool matches the task.
  2. Treat Plans as Portals: Use your planning document as a high-level interface to your codebase. Ask your AI to visualize and organize the relevant parts of the system to help you make informed decisions.
  3. Share Early: Stop working in silos. Share your plans with teammates to leverage their context and ensure alignment before the agent begins implementation.