Limitations of Agentic AI IDEs Like Cursor and Conductor

The author ditched Cursor (VS Code fork) and Conductor after heavy use, citing key pain points in agent workflows. Cursor excels for simple tasks but lacks robust multi-agent management. Conductor attaches GitHub repos to workspaces, spins up git worktrees, runs custom setup scripts (e.g., conductor-setup.sh for .env.local, seed data migration—ignored by git), and supports parallel agents with Opus/GPT Codex switching. It provides inline code reviews via a basic text diff viewer (like GitHub PR), comment integration into AI chats, and PR status polling (CI/CD, tests, Claude reviews).

However, Conductor's diff viewer is 'just a plain text viewer'—no LSP features like jump-to-definition or find-references. For complex changes (>3-4/10 complexity), the author constantly context-switched to Cursor instances (Conductor supports this natively), copying conversations over. This broke flow: 'I was always forced to open up another application when things got a little bit too complicated.' PR automation also needs 'babysitting' due to failures in CI/CD or merges.

Tradeoff: Agent tools prioritize AI code gen/review over engineer navigation, forcing hybrid setups that fragment workflows.

'Every now and then... when the code change got a little bit too crazy... I was still forced to open up Cursor or VS Code or true text editor.' (Highlights why basic viewers fail at scale, pushing for full editors.)

Replicating Agent Workflows in Neovim with Git Worktrees

To fix app-switching, the author rebuilt everything in Neovim (Kickstart.nvim config by TJ DeVries)—a 'barebones' setup with Vim keybindings (retained from VS Code/Cursor). Core: Custom 'yw' script (Claude-generated) mimics Conductor workspaces.

yw flow:

  1. yw test-workspace: Fetches main, creates branch/worktree.
  2. Runs conductor-setup.sh (.env, seed data, Vercel dir).
  3. npm install && npm run dev.
  4. Launches Claude Code CLI instance.

Neovim handles reviews: space gs (git status), space hd (horizontal diff), full LSP (gd=definition, grr=references). Added diffview.nvim plugin (space do) for VS Code-like file sidebar + side-by-side diffs—'literally the same exact thing.'

This scales to 7-8/10 complexity natively (vs Conductor's 3-4/10), using Neovim only as 'text reviewer' with LSP power. For 9-10/10, fallback to Cursor ('skill issue,' expects improvement).

'This is literally just a plain text viewer in Conductor... Now, within my new Neovim setup, I'm able to handle like a seven or eight out of 10 complex workflows.' (Quantifies LSP's edge for real engineering.)

Terminal Choice: Warp Over tmux/CMUX for AI CLI Integration

Initially used CMUX (tmux wrapper on Ghostty lib by HashiCorp co-founders)—vertical workspaces, horizontal tabs (Cmd+1-4 / Ctrl+1-4), first-class Claude/GPT Codex CLI with notifications (e.g., task complete alerts).

Switched to Warp (sponsor, but 'hear me out'): Better CLI detection (auto-senses Claude Code), notifications, multi-model support (Warp AI, but author prefers Claude/GPT Codex CLI). Key win: 'Finer UI/UX improvements' beyond pure terminal—e.g., progress tracking without full GUI bloat. Churned off Warp 2021-2025 for AI features, returned for Claude-first-class support.

Tradeoff: Warp > CMUX for polish; both beat iTerm2. No model preference lock-in.

'The one feature that really makes me like Warp over CMUX... it's not just a pure terminal and it has a little bit of like finer UI UX improvements.' (Explains terminal evolution for AI agents.)

Automating PRs and Scaling with Claude Code Skills

Replicates Conductor's PR panel via Claude Code CLI 'babysit-pr' skill + /loop (1-min cron):

  • gh CLI pulls PR.
  • Checks CI/CD statuses.
  • Auto-merges if green; fixes reds/comments or pings author.

Used for Yorby.ai (social media tool: viral DB + script writer). Claude as 'cheat sheet': 'Explain how to do XYZ in Neovim like VS Code.' Ramped up fast despite 5-year Vim hiatus.

When Cursor helps: Rare mega-diffs. Overall: Faster/flexible for 'real development work.'

'If it's all green and ready to go, then merge it automatically. If anything becomes red... ping me.' (Shows practical agent handoff limits.)

Key Takeaways

  • Use git worktrees + custom setup scripts (e.g., yw) for isolated AI agent workspaces—bypasses .gitignore for local data.
  • Neovim (Kickstart + diffview.nvim) + LSP crushes basic diff viewers for 70-80% of reviews; fallback to Cursor only for extremes.
  • Pick terminals by AI CLI support: Warp/CMUX for notifications/progress > plain tmux.
  • Build Claude skills like 'babysit-pr' + /loop for hands-off PRs, but expect babysitting.
  • Leverage AI (Claude) to learn/editor configs—'hey look at my neovim config and explain...'
  • Rank tools by complexity handled: Neovim (7-8/10) > Conductor (3-4/10) > plain viewers.
  • Retain Vim bindings in VS Code/Cursor for smooth transitions.
  • Rebuild agent flows in familiar editors before buying new IDEs—'this is not that special.'