Skip to content

Coding agents

You open a terminal, start Claude Code, Codex or Gemini CLI, give it a task, read the diff and approve. That is where most teams are in 2026. The tools below are that, plus a layer for running several agents at once.

This is the first of three levels. It works well, and it ends where your reading ends: every change still goes through you.

  • Claude Code (Anthropic). Terminal agent, also in VS Code, JetBrains, a desktop app and the web. Permission modes from approve-every-step to full auto. Agent teams of 3 to 5 sub-agents are the recommended ceiling.
  • Codex CLI (OpenAI), Gemini CLI (Google), Copilot CLI (GitHub), Cursor agent. Same shape: a task in, a diff out, you approve.

All of them stop at one task. A plan mode writes a plan for that task and the plan is gone when the task is done. Keeping the plan as files in the repo is what the spec-driven tools add.

Once one agent works, the next thing people do is open five. These apps give each agent its own workspace (usually a git worktree) and tell you which one needs you.

  • cmux. Open source macOS terminal built on Ghostty, with vertical tabs, split panes and a sidebar that shows each workspace’s branch, PR status and latest notification. A terminal first: it runs whatever you type. GPL-3.0; the server parts are BSL 1.1.
  • Conductor. Mac app that runs Claude Code, Codex and Cursor agents in parallel, each in its own workspace, and lets you review and merge. Closed source, paid plans.
  • Superset. Mac app (Linux experimental) that runs Claude Code, Codex and OpenCode in isolated worktrees, with a CLI, a TypeScript SDK, an MCP server and an iPhone app for checking in. Source-available (Elastic License 2.0), free tier.
  • Vibe Kanban. A kanban board where each card is a task for Claude Code, Codex, Gemini CLI or Amp, run in its own workspace with a branch, terminal and dev server; diff review, inline feedback, open and merge PRs. Apache-2.0. The README says the project is being sunset. It was bloop’s pivot away from COBOL modernisation.

All four run the agents on your machine. Five agents in five worktrees share one laptop, so a full test suite or an e2e run after each change is five full suites on the same CPU. In practice the slow check runs less often, and the diff is approved on the unit tests alone.

  • You are the bottleneck. Five agents make five diffs, and you read all of them. Anthropic’s own data says people accept 93 percent of permission prompts, which is approving, not reviewing. On a bounded task this shows within the hour: a repo with five independent issues and a 90 minute limit ends with two issues fixed well and the rest as notes, because finding the issues and reading each diff is where the time goes.
  • Nothing remembers. The plan lives in the chat. A second agent, or the same one tomorrow, does not know what was decided.
  • The check is you. Passing tests is taken as proof, and it is not: agents write the tests they then pass.

Each of these is what the next two levels try to fix: spec-driven tools keep the plan as files and check them; the software factory tries to take the person out of the reading loop.

Last updated: