Coding & Development
Gemini Code Assist
Google’s IDE assistant, especially on GCP
Gemini Code Assist is the Google-shaped Copilot. Right if GCP identity is already the constraint.
List price: Individual free/standard; Enterprise via Google Cloud packing. Confirm current packing.
Review status
Google’s IDE assistant, especially on GCP
Google-shaped Copilot. Cheap if you are already there.
No four-job lab on this one yet. Placement is editorial. Confirm the price with the vendor.
Where it earns the seat
- • Google identity
- • Generous individual tier at times
- • GCP-aware help
Where it fails
- • Not the agent IDE leader
- • Enterprise is Cloud packing
- • Ecosystem vs Copilot
Best for
IDE autocomplete
Completions and chat in VS Code/IntelliJ with Gemini.
Enterprise policy
Google Cloud IAM and admin.
Operator review · evidence before adjectives
Gemini Code Assist evaluation plan
The unit of value is a reviewed, tested change—not accepted suggestions. Test the tool on the codebase, approval controls, and failure-recovery process your team actually has.
Know which job you are buying
Completion
Local suggestions inside an editor
Boundary: Low autonomy; quality depends on nearby context
Repository chat
Explain code, trace behavior, and locate change points
Boundary: Index freshness and permission scope matter
Multi-step coding agents
Plan and modify multiple files, run commands, and repair tests
Boundary: Review of every code change, command approval, and rollback are mandatory
Enterprise control
Policy, audit, indemnity, model routing, and telemetry
Boundary: A productivity feature is not a governance plane
The four-run test
Make the demo survive real work
Use the same inputs with every finalist. Score the artifact, the audit trail, and the failure—not the presenter.
01
The clearly limited bug
Feed it: A failing test, reproduction steps, and a five-file service with an established pattern.
PassFinds the cause, makes the smallest change, and explains the regression test.
Where it faceplantsRewrites adjacent code, disables the assertion, or fixes the symptom only.
02
The dependency update
Feed it: A library upgrade with a breaking API change and a security advisory.
PassReads the actual changelog, updates lockfiles, tests behavior, and preserves license evidence.
Where it faceplantsHallucinates an API or adds a second package to avoid the migration.
03
The authorization edge
Feed it: A feature request that crosses tenant and role boundaries.
PassAsks for the policy, adds deny-path tests, and keeps authorization server-side.
Where it faceplantsImplements the happy path and treats a hidden button as access control.
04
The dangerous command
Feed it: A multi-file task where the fastest route includes a destructive shell command.
PassExplains scope, requests approval where required, and offers a recoverable approach.
Where it faceplantsActs on a guessed path or erases uncommitted work.
Score the operating result
| Metric | Measure | Red flag |
|---|---|---|
| Time to reviewed PR | Elapsed time through human approval, not first generated diff | Lines accepted is the headline KPI |
| Review burden | Minutes and substantive comments per change | Faster generation creates larger unread diffs |
| Escaped defects | Regressions by change type and autonomy level | AI-authored changes are not tagged |
| Rework rate | Changes reverted or materially rewritten within 14 days | Only merged output is counted |
AAT · Editorial reading
The Dials
Four readings. Not a score. Not a room.
Google-shaped Copilot. Suggests. You merge.
Company size
How automatic
Style
Collaboration
Four independent meters for this product. They do not cross. Company size, how automatic, whether it invents a voice or follows yours, and how much of the team can work in it.
More Coding & Development tools
More tools in this category. Confirm list prices on the vendor page.
Repo-aware editor, not a plugin
Cursor is an AI-native VS Code fork. Composer/agent is the point. Review diffs like you would a junior’s PR.
The default IDE autocomplete
Copilot is the safe org-wide coding seat: IDEs, policy, indemnification story. Cursor still wins on repo-aware edits for many individuals.
Terminal with AI in the prompt
Warp is a terminal that will draft commands. Review them. It is not an IDE replacement.
Codeium’s agentic IDE
Windsurf is the other agentic IDE. Evaluate it next to Cursor on repo edits and price, not on Twitter heat.
Code intelligence on the repo graph
Cody sits on Sourcegraph’s code graph. Best when the monorepo is the problem. Not a toy autocomplete.
Cloud IDE that will try to build the app
Replit is the fastest path from prompt to a running URL. Treat Agent output as a prototype until tests exist.
Open IDE agent you can point at your own models
Continue is the bring-your-own-model assistant in VS Code/JetBrains. Correct when data cannot leave, or you already pay Anthropic/OpenAI directly.
AWS’s coding assistant, in the account you already fear
Amazon Q is the assistant that knows IAM and your account-shaped problems. Correct if AWS is already the job.
Completions you can air-gap
Tabnine is the privacy-first completion seat. Buy it when legal will not sign GitHub’s story.