136 Ready-to-Use Codex Subagents:
The Ultimate Collection You Need to Bookmark
Stop writing the same prompts over and over. These plug-and-play agents are already configured, categorized, and ready to drop into your workflow.
·
⏱ 8 min read
·
🔖 Bookmark this
If you’ve been using OpenAI’s Codex, you’ve probably noticed the same frustration: the more you ask it to do at once, the blurrier the results get. Code review + documentation + debugging all in one context window? It starts losing focus fast.
That’s exactly why subagents exist — and why awesome-codex-subagents might be one of the most useful GitHub repos to land in the AI dev tools space this year.
TL;DR
awesome-codex-subagents is a free, open-source GitHub repo by VoltAgent with 136 pre-built Codex subagent definitions across 10 categories — backend, frontend, security, AI/ML, DevOps, business, and more. MIT licensed. Drop in a .toml file and you’re running.
Why One Agent Isn’t Enough
Think about how a real team works. You wouldn’t ask one person to simultaneously review your code for security vulnerabilities, rewrite the documentation, and write new tests. You’d split the work.
Subagents solve this for AI. Each one runs in its own isolated context window, focused on one job. This means:
- Less context bleed between tasks
- Sharper, more consistent output per domain
- Reusable configurations you and your team can share
- Parallelism — run multiple agents on the same task simultaneously
Codex introduced subagents as an official feature in March 2026. awesome-codex-subagents dropped almost immediately after — and it already has 319+ GitHub stars as of this writing.
How to Install (It Takes 2 Minutes)
No npm install. No config file maze. Just clone and copy.
# Step 1: Clone the repo
git clone https://github.com/VoltAgent/awesome-codex-subagents
# Step 2a: Install globally (available in all projects)
mkdir -p ~/.codex/agents
cp categories/04-quality-security/code-reviewer.toml ~/.codex/agents/
# Step 2b: Or install per-project (higher priority)
mkdir -p .codex/agents
cp categories/01-core-development/backend-developer.toml .codex/agents/
⚠️ Important: Explicit Delegation Required
Codex does not auto-spawn subagents. You need to call them by name in your prompt. Think of it like delegating to a team member — you have to assign the work explicitly.
Here’s what an explicit delegation prompt looks like in practice:
The 10 Categories (And Which Ones to Start With)
Here’s the full breakdown, with notes on the most useful agents in each category.
01
Core Development (12 agents)
Your everyday dev toolkit. Includes backend, frontend, fullstack, mobile, GraphQL, WebSocket, and more.
⭐ Start here: ui-fixer — makes the smallest safe patch for a reproduced UI bug. Saves you from overfixing.
02
Language Specialists (26 agents)
One agent per language/framework: Python, TypeScript, Rust, Go, Java, Ruby on Rails, Next.js, Flutter, Swift, Kotlin, PHP, Elixir, and more. Includes framework version specifics (e.g., Rails 8.1, Next.js 14+, Django 4+).
⭐ Start here: typescript-pro or python-pro — your most-used language, finally with its own context.
03
Infrastructure (16 agents)
Cloud, DevOps, Kubernetes, Docker, Terraform, Terragrunt, SRE, incident response, networking, and platform engineering.
⭐ Start here: incident-responder — gives you a structured runbook approach when something breaks at 2am.
04
Quality & Security (16 agents)
Code review, security audits, accessibility testing, performance engineering, QA automation, penetration testing, chaos engineering, and compliance.
⭐ Start here: reviewer — a PR-style review for correctness, security, and regressions. Pairs perfectly with security-auditor.
05
Data & AI (12 agents)
Data engineering, ML, MLOps, NLP, LLM architecture, PostgreSQL, prompt engineering, and data science.
⭐ Start here: llm-architect — useful if you’re building systems on top of language models, not just using them.
06
Developer Experience (13 agents)
Documentation, CLI tools, Git workflows, refactoring, legacy code modernization, build systems, dependency management, and Slack platform development.
⭐ Start here: legacy-modernizer — if you have old code that needs upgrading without breaking everything.
07
Specialized Domains (12 agents)
Blockchain/Web3, fintech, IoT, embedded systems, game development, payment integration, SEO, quant analysis, and M365 administration.
⭐ Start here: payment-integration — payment logic is error-prone enough to deserve its own focused agent.
08
Business & Product (11 agents)
Product manager, business analyst, UX researcher, technical writer, legal advisor, content marketer, customer success, scrum master, project manager, and WordPress expert.
⭐ Start here: ux-researcher — underrated. Forces structured user research framing before jumping into solutions.
09
Meta & Orchestration (11 agents)
Workflow orchestration, multi-agent coordination, task distribution, context management, error recovery, IT ops, and knowledge synthesis.
⭐ Start here: workflow-orchestrator — if you’re running complex multi-step workflows, this is your conductor agent.
10
Research & Analysis (7 agents)
Competitive analysis, market research, data research, documentation research, trend analysis, search, and synthesis.
⭐ Start here: docs-researcher — verifies framework APIs from documentation before your code depends on them. Prevents hallucination in API references.
Smart Model Routing: The Detail Most People Miss
Each .toml file includes a model field that routes the subagent to the right model automatically. According to the repo’s README, the routing works like this:
| Model | Use Case | Example Agents |
|---|---|---|
| gpt-5.4 | Deep reasoning tasks — architecture reviews, security audits, financial logic | security-auditor, architect-reviewer |
| gpt-5.3-codex-spark | Fast scanning, lightweight research, documentation lookup | search-specialist, docs-researcher |
Note: Model names are as specified in the repo’s .toml files. You can override them by editing the model field directly.
You can override any model assignment by editing the .toml file directly. This matters if you have API cost constraints or want to experiment with different inference speeds.
🛡️ Sandbox Modes: Built-in Safety Rails
Every agent has a sandbox_mode field that controls filesystem access:
• read-only — for reviewers and auditors. They analyze without touching files.
• workspace-write — for developers and engineers that need to create or modify files.
3 Workflows Worth Copying Right Now
These are pulled directly from the repo’s example prompts. Copy, adapt, and use.
🔍 Workflow 1: Parallel PR Review
“Review this branch with parallel subagents. Have reviewer look for correctness, security, and missing tests. Have docs_researcher verify the framework APIs this patch depends on. Wait for both and summarize findings with file references.”
🐛 Workflow 2: Bug Investigation
“Investigate the broken settings flow. Have code_mapper trace the owning code paths, browser_debugger reproduce the bug in the browser, and frontend_developer propose the smallest fix after the failure is understood. Wait for the read-heavy agents first, then continue.”
💡 Workflow 3: Refactoring with Research First
“Use search_specialist to locate the code related to payment retries, knowledge_synthesizer to summarize the current design, and refactoring_specialist to propose a minimal refactor plan. Return a concrete action list.”
🧠 Claude Code User? There’s a Separate Repo for You
VoltAgent also maintains awesome-claude-code-subagents, a parallel collection built specifically for Claude Code instead of Codex. Same structure, different target platform.
5 Tips for Getting the Most Out of These Agents
1. Start with a single agent, not the whole catalog
Install one agent that solves a real pain point you have today. Get comfortable with explicit delegation before running parallel workflows.
2. Project-level installs override global ones
If you put an agent in .codex/agents/ (project-level), it takes precedence over the same-named one in ~/.codex/agents/. Use this to customize per project without breaking your globals.
3. The .toml files are easy to edit
Don’t treat them as black boxes. Open the file, read the instructions.text field. You’ll understand what the agent prioritizes — and can tune it for your codebase.
4. Pair a reviewer with a developer, not just one or the other
The best results come from running reviewer in read-only mode first, then handing findings to a domain developer agent. Let the auditor audit; let the builder build.
5. Check the CONTRIBUTING.md if you want to add your own
The repo welcomes contributions. If you build a solid agent for your workflow, you can PR it in and it might make the collection. Open source at its best.
The Bottom Line
The gap between developers who give AI structure and those who just type prompts is widening fast. awesome-codex-subagents is a practical, free, MIT-licensed way to get on the right side of that gap today.
