Pick a coding agent by where it runs, then trial two on one real task

By vibe-coding-notes · 0 stars

Pick the surface you already work in (terminal, editor or GitHub), then choose between two tools there by running the same small task in both.

### Context

- Agents come in three shapes, and most products ship more than one:
  - Terminal: runs in your shell on a local checkout (Claude Code, Codex CLI, Gemini CLI, GitHub Copilot CLI).
  - Editor: runs in an IDE with inline diffs (Cursor, VS Code with Copilot, Windsurf, JetBrains AI).
  - Cloud: runs in a remote VM or container and returns a branch or PR (Copilot cloud agent, Cursor Cloud Agents, Codex cloud, Claude Code on the web).
- On hard tasks the model matters more than the wrapper. The wrapper decides safety defaults, context handling and how easy the output is to review.
- All of them make the same classes of mistakes; none removes the need to verify.

### Options

| Tool | Surfaces | Instruction file | Isolation / safety | Strong at | Watch out for |
|------|----------|------------------|--------------------|-----------|---------------|
| Claude Code (Anthropic) | CLI, VS Code and JetBrains extensions, desktop app, web | `CLAUDE.md` (also reads `AGENTS.md`), `.claude/rules/` | Permission modes (manual, acceptEdits, plan, auto, bypass), hooks, built-in `--worktree` | Long multi-step terminal tasks, planning, subagents | Auto modes: review what ran |
| Cursor | Desktop editor (VS Code fork), Cloud Agents | `.cursor/rules/*.mdc`, `AGENTS.md` | Per-action approval in editor; cloud agents in isolated VMs | Tight edit-review loop in the IDE | Rules files multiply; cloud agents need repo access |
| OpenAI Codex | CLI (open source), IDE extension, cloud, desktop app | `AGENTS.md`; config `~/.codex/config.toml` | Sandbox modes (read-only, workspace-write, danger-full-access) plus approval policy; sandbox network off by default | Sandboxed local runs, parallel cloud tasks | Sandbox and approvals are two separate knobs |
| GitHub Copilot cloud agent | GitHub issues and agents panel, VS Code, `@copilot` in PRs | `.github/copilot-instructions.md`, `.github/instructions/*.instructions.md`, `AGENTS.md`, `CLAUDE.md` | Ephemeral GitHub Actions environment; admins must enable it | Issue to PR inside GitHub | Per-task time limit; branch protection can block it |
| Gemini CLI (Google) | CLI, open source | `GEMINI.md` (global in `~/.gemini/`) | Approval prompts, optional sandboxing | Large context, free tier for individuals | Fewer team-management features |
| Others (Windsurf, Cline, Aider, Zed, JetBrains Junie) | Varies | Many read `AGENTS.md` | Varies | Specific editor or model preferences | Check maturity and maintenance |

### Decision

- Pick a terminal agent when your workflow is driven by tests and CLI commands.
- Pick an editor agent when you want to see and accept each change as it happens.
- Pick a cloud agent for well-scoped, independent tickets you would hand to a junior developer.
- Pick by trial, not by blog comparison: run one bug with a failing test in two tools and compare whether each planned first, ran the tests, produced a readable diff, and how many approvals it needed.

### Consequences

- Good: `AGENTS.md` is read by most of these tools and stewarded by the Agentic AI Foundation (Linux Foundation), so switching later costs little if your instructions live there.
- Bad: pricing, limits and product names change monthly. Recent renames: Cursor "Background Agents" became "Cloud Agents"; GitHub Copilot "coding agent" became "cloud agent" (April 2026). Check the vendor page on decision day.
- Bad: connecting a private repo sends code to a third party. Check your organization's data policy first.

As of September 2026.

### Sources

- https://code.claude.com/docs/en/permission-modes
- https://cursor.com/docs/rules
- https://cursor.com/docs/cloud-agent
- https://learn.chatgpt.com/docs/codex/cli
- https://learn.chatgpt.com/docs/agent-approvals-security
- https://docs.github.com/en/copilot/concepts/agents/coding-agent/about-coding-agent
- https://github.blog/changelog/2026-04-01-research-plan-and-code-with-copilot-cloud-agent/
- https://geminicli.com/docs/cli/gemini-md/
- https://agents.md/