Суб-агенты в Codex App: команда AI-разработчиков

Codex App from OpenAI is not just a chat with a model, but a real command center for agent coding**. With the release of sub-agents in 2026, it allows the master agent to spawn specialized assistants who work in parallel in isolated contexts. One prompt – and instead of performing tasks consistently, you get multiple AIs running simultaneously, like a real dev team.

It fits perfectly into the philosophy of *wibcoding: you describe the idea, and Codex breaks down the task, delegates and collects the result. Sub-agents solve the main pain of big projects – pollution of the context and long expectations.

## How sub-agents work in Codex App

Codex does not automatically spawn sub-agents (this is OpenAI’s conscious decision to keep the lead agent focused). They will only be launched at your explicit request.

Mechanism:

• The main codex analyzes the problem.
• It generates up to 6+ sub-agents (depending on release).
• Each sub-agent receives its own isolated context, model, tools, and instructions.
• Sub-agents work in parallel in separate worktrees or cloud environments.
• The results are collected in a single response from the chief agent.

Key Terms (from official developer documentation.openai.com/codex):

• Subagent workflow - parallel running of specialized agents + results aggregation.
• *Custom Agents - You define the model, prompt and tools for the sub-agent.

Visibility: Codex App (macOS/Windows) and CLI are already available. In the VS Code extension - soon.

## How to run sub-agents: simple commands and prompts

In Codex App (graphic interface):

1. Open the project.
2. In the “Agents” sidebar or via chat, write a request.
3. An example of a prompt: Break the task into parallel sub-agents: one examines the code base, the second writes front-end components, the third tests and security audit. Use separate worktrees.”

B CLI (terminal):

• /agent list shows active agents.
• /agent switch <имя> – switch between them.
• Just say, "Spawn 3 sub-agents: Scout for exploration, Worker for implementation, Reviewer for code review."

AGENTS.md is your main configuration file. Codex reads it automatically before any job. Example of content:

Custom sub-agents
Reviewer: Always checks for bugs, security, and architecture compliance. Model: gpt-5.4-high
Scout: only reads files and makes a summary, unchanged.
Tester: runs tests and offers fixes.

## Real examples of use

Example 1: Large codebase research (bootstrap-repo) “Understand the entire repository and create comprehensive documentation.” → The master agent will save 5-10 subagents (one per module: auth, UI, API, DB). The result: in 10-15 minutes you have a full architecture.md + llms.txt. Speed: 2.5-3 times faster than a single agent.

**Example 2: Implementation of complex features (Telegram-bot with payments and AI) ** Prompt: “Implement the subscription feature. Spawn sub-agents:

• Frontend-agent - UI and Adaptive
• Backend-agent - route, Stripe, queues
• Test-agent - unit + integration tests
• Security-agent – key check and rate-limit Sub-agents work in parallel worktrees, the main collects and greases. A real case from the vibe-coding community: a full bot in one evening instead of 2-3 days.

**Example 3: Daily Workflow Review + Refactor ** In the morning: “Run in parallel:

1. Reviewer-subagent - code review of all PR
2. Refactor-subagent – Improve performance in Module X
3. Docs-Subagent – Update the Documentation All in the background while you're drinking coffee.

Example 4: Security audit of a large project “Spawn 4 sub-agents: one on secrets scanning, one on dependency vulnerabilities, one on SQL injection check, one on report.” Codex App shows real-time progress in sidebar.

## Useful chips and best practices for vibcoding

1. Always use AGENTS.md is your “contract” with Codex. Add permanent reviewers, testers and scout agents.
2. Select models for the task: the main one is gpt-5.4-high (reasoning), sub-agents are lightweight versions for speed.
3. Avoid Context Pollution: Sub-agents work in isolation → The main agent stays clean and smart.
4. Combine with memory-bank (from vibe-coding guides): put game-design-document.md, implementation-plan.md - sub-agents read them too.
5. Monitoring in App: The sidebar shows all agents, their status, token consumption and logs.
6. Token savings: One large prompt with sub-agents is often cheaper than 5 individual chat rooms.
7. **Security: Sub-agents work in a sandbox (local or cloud). Never give them access to secrets without explicit permission.
8. CLI + App Hybrid: Run heavy tasks in CLI (SSH, remote server), and monitor in a beautiful App.

## Pluses

• ** Speed: weeks of work → days (parallelism gives ×2.5-3).
• **Quality: Each sub-agent is a specialist (there is no “jun effect” of one large agent).
• **Scale: It’s easy to build whole apps, bots, websites without burnout.
• Going from Jun to Pro: You don’t write code anymore, you ** orchestrate an AI team.

Codex App with sub-agents is the next level after the simple “ask to write a feature”. Now you literally run an AI team like a techlid.