A terminal agent can plan, edit files, and run commands across a real repo. That power is the point, and it is also the blast radius. Start on a feature branch, keep the goal small, prefer a plan first, and let a human merge.
At 4:40 p.m. someone pastes last quarter’s install notes for “Gemini CLI” into a laptop that still has production credentials in the default shell path. The binary name on the blog is already outdated. The agent offers to clean up a folder. Nobody opened a throwaway branch first.
Terminal agents feel like magic because they can plan, edit, and run commands across a repo. That is also why careful teams stay nervous. Gemini CLI and Antigravity CLI sit in this family: more autonomy than inline complete, more blast radius than a side-panel sentence rewrite.
Surfaces by blast radius
| Surface | Autonomy | Best for |
|---|---|---|
| Gemini app chat | Low | Design talk, snippets, non-repo questions |
| IDE assist | Low-medium | Inline complete, small edits |
| Terminal CLI agent | Medium-high | Multi-step repo tasks with review |
| Antigravity-style platform | High | Agent-first workflows when the team is ready |
Chat is for talking. IDE assist is for small edits next to your files. A terminal agent can touch many files and run commands. An Antigravity-style platform leans even further into agent-first work. Same model family name does not mean the same review habit.
Safe starter kit
- Create a feature branch.
- State a small, testable goal.
- Prefer plan or dry-run modes when available.
- Watch file touches; stop if scope creeps.
- Run tests; open a PR; a human merges.
# Internal standard stub
# Approved terminal agent: (Antigravity CLI / Gemini CLI / none)
# Auth: work account / Cloud project
# Data: no customer PII in prompts
# Git: feature branch required
# Review: human required for auth, payments, migrations
# Last verified: YYYY-MM-DDMigration without panic
When binaries rename, inventory aliases, CI jobs, and wiki pages the same day. Re-auth, re-test one real workflow, update the standard stub. Dated docs beat tribal memory. Mid-2026 consumer transitions pushed many people from Gemini CLI toward Antigravity CLI; enterprise paths often differed. Treat names as map pins and re-check migration docs the week you train the team.
Guardrails that belong in the stub
- No production credentials in agent environments for experiments
- No force-push culture
- Cap minutes and cost where possible
- Log model and tool versions if audit matters
Practice this week
- On a throwaway repo, run one explain-failing-test style task.
- Capture the plan and the final diff.
- Write three rules your team would require before production use.
Common mistakes
- Copy-pasting pre-transition install notes
- Assuming enterprise and consumer timelines match
- Letting agents own merge buttons
- Training juniors on full autonomy first
What good looks like
A platform team allowed CLI agents only after a documented review checklist and a ban on production secret access. Incidents dropped. A startup let agents loose on main and spent a week restoring files. The tool was not the villain. Process was.
Ship one small, reviewed change this week using only these habits. Coding AI skills stick when they produce merges you can defend, not when they produce impressive demos you cannot explain.
Quick recap
- Terminal agents are multi-step tools with real blast radius.
- Branch, scope, plan, human merge.
- Next: the ask-edit-test-review loop in detail.
Series notes
This is Part 2 of the Gemini coding surfaces tutorial (GM23). Previous: IDE assist vs chat. Next: the ask-edit-test-review loop.
Sources
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
