Why I Choose Codex Over Other AI Coding Tools: Instant Interrupt & Parallel Agents

After two years of using various AI coding tools, the author settled on Codex because it delivers practical time savings in daily development work, not just chat-style assistance inside an IDE.

Instant Interrupt: Correct Course Mid‑Task

With other agents, a wrong turn (e.g., refactoring auth with sessions instead of JWT) forced the user to watch the agent finish the entire wrong step, then manually roll back. Codex’s instant_interrupt setting lets you type a single correction — “Stop, use JWT not session” — and the agent immediately replans without waiting for the current step to finish. This eliminates the “wait a minute + roll back many files” cycle.

Parallel Agents via git worktree

Previously, tasks like implementing payment, writing tests, and updating docs had to be done serially. Using git worktree, the author creates three isolated working directories, each on its own branch, and launches a Codex agent in each:

# Check out three independent workspaces in the same repo, each on its own branch
git worktree add ../app-pay -b feat/payment
git worktree add ../app-test -b feat/payment-tests
git worktree add ../app-docs -b feat/payment-docs

# Start one agent per terminal, all working simultaneously
cd ../app-pay && codex "Implement payment integration per docs/payment.md"
cd ../app-test && codex "Add unit tests for the payment module"
cd ../app-docs && codex "Update payment‑related documentation"

Each agent modifies its own files and runs its own tests without conflicts. The /agents panel shows all sessions and allows instant switching. Three serial tasks become parallel, saving significant time.

Persistent Project Rules with AGENTS.md

Repeated instructions — use pnpm not npm, test command format, lint requirements — are written once in AGENTS.md at the repository root. Codex reads this file every session, so rules are never forgotten even during long‑running tasks. The author observed a noticeable increase in first‑run success rates and less time spent searching package.json for commands.

<!-- AGENTS.md at repo root, shared by the team -->
# Project Conventions
## Build and Test
- Use pnpm install to install dependencies, not npm, to avoid lockfile corruption
- Run unit tests: `pnpm test --filter <package-name>`

## Code Style
- All TypeScript, no `any` allowed
- New APIs must use zod schema for validation

Cloud Execution: Work Continues After Laptop Close

Codex Cloud runs tasks in remote containers. A large repo‑wide refactor can be started in the evening; results are ready the next morning. Progress is viewable from a phone. This removes the “idle waiting” period entirely.

Safety by Default: Sandbox & Approval Policy

Codex runs in a sandbox ( sandbox_mode = "workspace-write") that restricts file operations to the workspace. The approval policy ( approval_policy = "on-request") only pauses for dangerous commands like rm -rf or git push. The author’s config:

# ~/.codex/config.toml
# Sandbox: only operate inside current workspace
sandbox_mode = "workspace-write"

# Approval: ask only for dangerous ops (rm -rf, git push, etc.)
approval_policy = "on-request"

# Allow new input to interrupt current operation immediately
instant_interrupt = true

This balances automation (read code, run tests, edit files) with safety.

Cost‑Effective Model Strategy

The default model gpt-6.1-sol offers near‑flagship capability at one‑tenth the cached‑input cost, sufficient for daily feature work, refactors, and test writing. For hard problems requiring cross‑session context or long reasoning, the author temporarily switches to a higher model via the command line:

# --model overrides config; use when needed without editing config
codex --model gpt-6-astra-medium "Audit the entire order module and find concurrency issues"

Daily cheap model, occasional upgrade — keeps costs controllable.

Caveats

Don’t max out reasoning intensity; medium is usually enough. Max is slower and more expensive. gpt-6.1-sol cannot disable reasoning; setting none or minimal in config is ignored. AGENTS.md has a 32 KiB limit; put critical content first.

MCP tools bypass Codex’s sandbox; their server security must be managed separately.

Recommendation

If you only need autocomplete, existing IDE plugins are enough. But if you want AI to own a whole task — read code, edit, run tests, verify pass/fail — while you intervene at any point to change direction, Codex is the most seamless tool the author has used. Setup: write a clear AGENTS.md and enable instant_interrupt. After that, there’s no going back.