The AI coding tool map, in five layers
The category has enough products that the useful question is no longer which tool but which layer.
AI Tools Hub Editorial · published · updated
Five layers
| Layer | What it does | Examples |
|---|---|---|
| Completion | Suggests the rest as you type | Supermaven, GitHub Copilot, Codeium |
| Editor agent | Edits across files on instruction | Cursor, Windsurf, Zed, Augment Code |
| Autonomous agent | Takes an issue, returns a result | Devin, Replit Agent, Bolt, Lovable |
| Review | Reads the PR and comments | CodeRabbit, Greptile, Qodo |
| PR operations | Keeps PRs small, manages merging | Graphite |
Return by layer
Completion has the highest usage frequency and near-zero learning cost, which makes it the least likely purchase to go unused. Editor agents vary enormously by individual; trial with two or three people before buying seats for everyone.
Autonomous agents work on clearly bounded, repetitive changes and disappoint on anything involving design judgement. Review pays off where PRs queue; PR operations only pays off where PRs are large and slow.
Avoid overlap
Teams routinely pay for three editors at once, which nobody can use simultaneously, and two review tools, which produces two sets of automated comments that everyone learns to ignore.
A sanity check on spend
A five-person team covering completion, editor agent and review typically lands between $200 and $400 a month, or $40–$80 per developer. Materially above that, look for a subscription nobody opens.
FAQ
Where should we start?
Completion. Highest frequency, lowest learning cost, least likely to be abandoned.