VoltAgent awesome-codex-subagents — Summary

VoltAgent awesome-codex-subagents is the OpenAI Codex CLI counterpart to awesome-claude-code-subagents, providing 166+ subagents in TOML format across 13 categories. The TOML format is Codex-native with fields like model, model_reasoning_effort, sandbox_mode, and developer_instructions — distinct from Claude Code's YAML frontmatter + markdown body format. Three additional categories beyond the 10 in the Claude version cover AI governance/safety (5 agents), platform engineering/IDP (5 agents), and LLMOps/evals/observability (5 agents). Smart model routing assigns GPT-5.4 to deep-reasoning tasks (architecture reviews, security audits, financial logic) and GPT-5.3-codex-spark to fast scanning/synthesis. Sandbox mode philosophy distinguishes read-only agents (reviewers, auditors) from workspace-write agents (developers, engineers) at the TOML level — a security-by-default pattern absent from the Claude Code version. Compared to voltagent-subagents, this repo is architecturally identical in catalog philosophy but uses TOML format instead of Markdown and targets a different runtime (Codex CLI instead of Claude Code). The relationship between these two VoltAgent repos is the most explicit example of dual-harness mirroring in this batch.

VoltAgent awesome-codex-subagents — Overview

Origin

Same maintainer as awesome-claude-code-subagents (VoltAgent). 4,894 stars, last push 2026-05-25. MIT license. "Written specifically for Codex and aligned with the official docs."

Philosophy

From README:

"This repository serves as the definitive collection of Codex Subagents, specialized AI assistants designed for specific development tasks. Written specifically for Codex and aligned with the official docs."

Key differences from Claude Code version:

1. TOML format (Codex-native) vs Markdown
2. sandbox_mode field — explicit security boundary per agent
3. model_reasoning_effort — controls reasoning depth
4. developer_instructions replaces markdown body
5. GPT model routing instead of Claude model routing

TOML Format Philosophy

From README:

"Each subagent includes a model field that automatically routes it to the right model — balancing quality and cost."

Sandbox mode:

"Read-only agents (reviewers, auditors): sandbox_mode = 'read-only' — analyze without modifying. Workspace-write agents (developers, engineers): sandbox_mode = 'workspace-write' — create and modify files."

Installation Model

More manual than Claude Code version:

mkdir -p ~/.codex/agents
cp categories/01-core-development/backend-developer.toml ~/.codex/agents/

No interactive installer, no plugin marketplace integration. Codex does NOT auto-spawn custom subagents — requires explicit delegation in prompts.

Three Additional Categories

Categories 11-13 extend beyond the 10 in the Claude Code version:

• 11-ai-governance-safety — AI safety and governance specialists
• 12-platform-engineering-idp — Internal Developer Platform engineers
• 13-llmops-evals-observability — LLMOps, evaluation, and observability specialists

VoltAgent awesome-codex-subagents — Architecture

Distribution

Manual copy. No plugin marketplace, no installer script.

# Global agents
mkdir -p ~/.codex/agents
cp categories/01-core-development/backend-developer.toml ~/.codex/agents/

# Project-specific
mkdir -p .codex/agents
cp categories/04-quality-security/reviewer.toml .codex/agents/

Directory Structure

categories/
  01-core-development/         # 12 agents (.toml)
  02-language-specialists/     # 31 agents (.toml)
  03-infrastructure/           # 16 agents (.toml)
  04-quality-security/         # 19 agents (.toml)
  05-data-ai/                  # 13 agents (.toml)
  06-developer-experience/     # 15 agents (.toml)
  07-specialized-domains/      # 13 agents (.toml)
  08-business-product/         # 16 agents (.toml)
  09-meta-orchestration/       # 11 agents (.toml)
  10-research-analysis/        # 9 agents (.toml)
  11-ai-governance-safety/     # 4 agents (.toml)  [NEW vs Claude version]
  12-platform-engineering-idp/ # 4 agents (.toml)  [NEW]
  13-llmops-evals-observability/ # 4 agents (.toml) [NEW]

Agent Format (TOML)

name = "backend-developer"
description = "Use when a task needs scoped backend implementation..."
model = "gpt-5.4"
model_reasoning_effort = "high"
sandbox_mode = "workspace-write"
developer_instructions = """
Own backend changes as production behavior...
"""

Storage Locations

Project-specific overrides global on naming conflict.

Target AI Tools

OpenAI Codex CLI only.

Required Runtime

Codex CLI (OpenAI).

VoltAgent awesome-codex-subagents — Components

Totals

Category Breakdown (166 agents across 13 categories)

Three New Categories (vs Claude Code version)

11 AI Governance & Safety

Agents covering AI policy compliance, model safety auditing, AI ethics, bias detection, and governance frameworks.

12 Platform Engineering / IDP

Internal Developer Platform specialists: platform tooling, self-service infrastructure, developer portals, IDP architecture.

13 LLMOps, Evals & Observability

LLM operations specialists: model evaluation pipelines, observability instrumentation, evals framework design, monitoring for AI systems.

Model Routing

Sandbox Modes

Sample TOML (backend-developer)

name = "backend-developer"
description = "Use when a task needs scoped backend implementation..."
model = "gpt-5.4"
model_reasoning_effort = "high"
sandbox_mode = "workspace-write"
developer_instructions = """
Own backend changes as production behavior with explicit data, auth, and failure-path integrity.

Working mode:
1. Map entry point, domain logic boundary, and persistence side effects.
2. Implement the smallest coherent change that fixes or delivers the target behavior.
3. Validate behavior under normal and high-risk failure paths.

Focus on:
- request/event entry points and service boundary ownership
- input validation and contract-safe output behavior
...

Quality checks:
- validate one critical success path and one high-risk failure path
"""

Representative Agents

• Category 01: api-designer, code-mapper (NEW), ui-fixer (NEW), backend-developer, frontend-developer
• Category 02: typescript-pro, python-pro, golang-pro, rust-engineer
• Category 03: kubernetes-specialist, terraform-engineer, cloud-architect
• Category 04: security-auditor, code-reviewer, qa-expert
• Category 09: agent-installer, workflow-orchestrator, multi-agent-coordinator

VoltAgent awesome-codex-subagents — Uniqueness & Positioning

Differs From Seeds

No seed targets Codex CLI as primary tool — this is the only Codex-native catalog in this batch. The TOML format with sandbox_mode and model_reasoning_effort fields has no equivalent in any seed framework. The dual-repo strategy (pairing with voltagent-subagents for Claude Code) represents an explicit "one content, two harnesses" philosophy implemented through separate repos rather than an adapter framework.

Relationship to Sister Repo

The most important thing to understand about this repo is its relationship to voltagent-subagents: same catalog structure (categories 1-10 are identical domains), same agent names, same organizing company — but different format, different runtime, and three additional categories (11-13). This is the most explicit same-company dual-harness strategy in this batch.

Sandbox Mode Innovation

The TOML sandbox_mode = "read-only" for reviewers/auditors and sandbox_mode = "workspace-write" for developers/engineers is a security-by-default pattern that Claude Code's format cannot express. This is the most distinctive technical feature of this repo vs any other in the batch.

Observable Failure Modes

1. No auto-spawn: Users migrating from Claude Code expect auto-activation; Codex requires explicit delegation.
2. Model name drift: GPT model versions will be deprecated; TOML files with hardcoded model names will break.
3. Manual sync: No automated mechanism to keep this repo in sync with the Claude Code version — categories 11-13 only exist here; future Claude Code additions may not propagate.

Explicit Antipatterns

None stated. README focuses on "written specifically for Codex" positioning.