Google Agents CLI vs Codex CLI: Two Visions of Agent Development from the Terminal

agents-cligoogle-cloudadkcodex-cliagent-platformdeploymententerprise

Sketchnote diagram for: Google Agents CLI vs Codex CLI: Two Visions of Agent Development from the Terminal

Published: 2026-05-20. Sources: Google Developers Blog, GitHub google/agents-cli, Google Cloud I/O agent developer blog, InfoQ coverage, OpenAI Codex CLI docs, OpenAI Python SDK docs, OpenAI Plugins docs, OpenAI Skills docs.

The Core Distinction

This comparison hinges on a single architectural insight: Codex CLI IS the coding agent; Google Agents CLI builds agents ON Google Cloud. They occupy different layers of the stack, and understanding where they overlap and where they complement each other is essential for teams evaluating both.

  • Codex CLI — a terminal-native coding agent that writes code, runs tools, manages PRs, and orchestrates subagents. It is the agent.
  • Google Agents CLI — a scaffold/eval/deploy lifecycle tool for the Agent Development Kit (ADK). It is the factory line that coding agents drive to ship agents to Google Cloud.

Architecture Comparison

Dimension Codex CLI Google Agents CLI
Role Coding agent (writes, tests, deploys code) Agent lifecycle tool (scaffolds, evaluates, deploys agents)
What it produces Code changes, PRs, automated workflows ADK agent projects deployed to Google Cloud
Runtime Local terminal with kernel-level sandbox No runtime — invoked by a coding agent or human
Model GPT-5.5, codex-mini, o3-pro (OpenAI models) Model-agnostic (used by any coding agent)
Language TypeScript (CLI), Python (SDK) Python, TypeScript, Go, Java (via ADK)
Installation npm install -g @openai/codex uvx google-agents-cli setup or npx skills add google/agents-cli

How They Work Together

Google Agents CLI was explicitly designed to work WITH Codex CLI, not replace it. The integration path:

  1. Codex CLI handles the coding loop — writing agent code, implementing tools, debugging
  2. Agents CLI handles the deployment loop — scaffolding ADK projects, running evaluations, deploying to Google Cloud
  3. Agents CLI ships as skills (Markdown skill files) that teach Codex CLI how to invoke its commands

In practice, you tell Codex CLI:

  • “Scaffold a new ADK agent” → Codex invokes agents new
  • “Run evaluations against the eval set” → Codex invokes agents eval run
  • “Deploy this agent to Cloud Run” → Codex invokes agents deploy

Google Agents CLI: Core Workflow

Seven Bundled Skills

Agents CLI packages seven specialised skills covering the full agent lifecycle:

  1. Workflow — development lifecycle rules and code preservation
  2. ADK Code — Python API patterns for agents, tools, and orchestration
  3. Scaffolding — agents new project creation and enhancement
  4. Evaluation — agents eval run, agents eval compare, LLM-as-judge scoring
  5. Deployment — Cloud Run, GKE, CI/CD pipeline automation
  6. Publishing — Gemini Enterprise registration and distribution
  7. Observability — Cloud Trace integration and structured logging

Dual Operating Modes

  • Agent Mode — AI assistants execute CLI commands autonomously
  • Human Mode — developers run commands directly for manual oversight

Deployment Targets

  • Google Cloud Run (managed containers)
  • Google Kubernetes Engine (GKE)
  • Agent Engine (managed agent hosting)
  • Gemini Enterprise (enterprise distribution)

Codex CLI: Agent Development Story

Codex CLI’s own agent development capabilities have expanded significantly:

openai-codex Python SDK

The openai-codex / openai_codex package (migrated in v0.131.0) provides:

  • App-server JSON-RPC protocol for multi-turn agent threads
  • Concurrent turn routing
  • Approval mode APIs
  • Pinned runtime-generated types

codex exec

Non-interactive execution for CI/CD pipelines:

  • --output-schema for structured JSON output
  • Sandbox isolation for safe automation
  • Integration with GitHub Actions and CI runners

Plugin Marketplace

v0.131.0 introduced marketplace CLI commands:

  • Version-aware plugin sharing
  • Share checkout workflows
  • Default-enabled plugin hooks
  • Shared-workspace buckets

Skills Directory

SKILL.md files with YAML frontmatter:

  • Discoverable via unified mentions (@)
  • Composable with hooks and plugins
  • Community-shareable

Where They Compete

Despite the complementary positioning, competition exists at the platform level:

1. Agent Deployment Surface

  • Agents CLI pulls deployment toward Google Cloud (Agent Platform, Cloud Run, Gemini Enterprise)
  • Codex CLI pulls toward OpenAI infrastructure (Codex Cloud, cloud VM tasks, codex exec)
  • Teams must choose where their agents run in production

2. Evaluation Frameworks

  • Agents CLI provides agents eval run and agents eval compare with ground-truth datasets
  • Codex CLI relies on custom hooks, test suites, and CI/CD integration for evaluation
  • Google’s eval framework is more opinionated; Codex’s is more flexible

3. Enterprise Governance

  • Agents CLI leverages Google Cloud IAM, VPC Service Controls, and Agent Platform governance
  • Codex CLI provides enterprise config layers, approval policies, sandbox modes, and the Dell AI Factory on-premises option
  • Different governance models for different compliance requirements

4. Skill/Plugin Ecosystems

  • Agents CLI ships skills that teach coding agents Google Cloud patterns
  • Codex CLI has its own skills directory and plugin marketplace
  • The ecosystems don’t interoperate — skills are agent-specific

When They Complement

The strongest use case is combining both:

Developer → Codex CLI (write agent code)
              ↓
         Agents CLI skills (scaffold ADK project)
              ↓
         Agents CLI eval (validate against eval sets)
              ↓
         Agents CLI deploy (ship to Google Cloud)

This gives you Codex CLI’s coding intelligence with Google Cloud’s deployment infrastructure.

Decision Framework

If you need… Use…
A coding agent to write and test code Codex CLI
To deploy agents to Google Cloud Agents CLI (via Codex CLI)
To deploy agents to OpenAI infrastructure Codex Cloud / codex exec
End-to-end agent lifecycle on Google Cloud Both together
On-premises agent deployment Codex CLI + Dell AI Factory
Multi-cloud agent deployment Codex CLI + Agents CLI + provider-specific tools

Enterprise Deployment Comparison

Aspect Codex CLI + OpenAI Agents CLI + Google Cloud
Compute Codex Cloud VMs, local sandbox Cloud Run, GKE, Agent Engine
Governance Enterprise config layers, approval policies IAM, VPC Service Controls
On-premises Dell AI Factory partnership Anthos / GKE Enterprise
Model lock-in OpenAI models only Model-agnostic (ADK)
Billing OpenAI API usage Google Cloud billing
CI/CD codex exec, GitHub Actions agents deploy, Cloud Build

Practical Recommendation

For Daniel’s agentic pod architecture:

  1. Use Codex CLI as the primary coding agent — it handles the inner loop of code generation, testing, and PR management
  2. Add Agents CLI skills when deploying to Google Cloud — a dedicated “deployer” agent role in the pod can use Agents CLI for all GCP deployment tasks
  3. Keep deployment-surface flexibility — don’t lock into a single cloud provider’s agent hosting
  4. Watch the ecosystem convergence — as Google expands A2A (Agent-to-Agent) protocol integration and OpenAI expands its plugin marketplace, the interoperability story will evolve

What Changed at Google I/O 2026

The I/O 2026 announcements (19 May) strengthened Agents CLI’s position:

  • Antigravity 2.0 now supports dynamic subagents and scheduled tasks, giving Google a more complete agent platform
  • Managed Agents in Gemini API provides agent-as-a-service execution, competing with codex exec
  • Android CLI 1.0 reached stability, showing Google’s commitment to CLI-based agent tooling
  • Agents CLI gained broader visibility as the unifying deployment tool across Google’s agent ecosystem

The key competitive nuance remains: Google Agents CLI is designed to work WITH Codex CLI, but it competes at the platform level by pulling agent deployment toward Google Cloud versus OpenAI’s own infrastructure.