Pick an AI coding agent by where your code lives, how much autonomy you want, and what you can afford to send to a vendor — not by leaderboard screenshots.
The three jobs people confuse
- Inline completion — finish the line you are typing (Copilot-class).
- IDE chat / edits — explain a file, apply a multi-file patch you review (Cursor-class).
- Autonomous agent — run commands, open PRs, iterate on failing tests (open-source agents, some cloud agents).
Most “which is best?” posts mix these. Decide the job first.
Decision table
| Need | Lean toward | Why |
|---|---|---|
| Stay inside VS Code / JetBrains with low setup | GitHub Copilot | Mature completion + chat; org SSO common |
| Agent-style edits in a dedicated editor | Cursor | Repo-aware chat, apply/reject diffs, rules |
| Code never leaves your network | Open-source local stack (Ollama + Continue / Aider / Cline) | You run the model; quality depends on hardware |
| Cheap experiments on public repos | Free tiers + open agents | Expect weaker long-context than paid frontier models |
Browse the full set under Code / Agents.
Practical checklist (10 minutes)
- Privacy: Can diffs and filenames leave the machine? If no → local / self-hosted only.
- Repo size: Monorepos need strong context and ignore rules; weak agents thrash.
- Review culture: If every change must be human-reviewed, prefer IDE chat over unattended agents.
- Budget: Count seats × tokens, not just the sticker price.
- Escape hatch: Can you export rules/prompts and switch tools in a week?
What we do not recommend
- Buying a tool because a Twitter thread claimed “10×.”
- Pointing an autonomous agent at production credentials.
- Expecting one agent to replace design review, security review, and product taste.
Related on AI Tools Nav
- Category: Code / Agents
- Category: Open-source / Self-host
- Items: Cursor, GitHub Copilot, Ollama
Editorial. We list tools we can verify; we do not sell rankings.