GitHub Agent HQ Model Selection: Choosing GPT-5.4 vs GPT-5.3-Codex, Custom Agents, and the Multi-Agent GitHub Workflow

GitHub’s Agent HQ — the platform that lets developers assign coding tasks to Copilot, Claude, and Codex directly from issues, pull requests, and the Agents tab — gained model selection for third-party agents on 14 April 2026¹. Previously, the Codex agent on GitHub ran with a fixed model. Now you can choose between GPT-5.2-Codex, GPT-5.3-Codex, and GPT-5.4 when kicking off a task¹. This article covers how model selection works, when to pick each model, how custom .agent.md files extend the system, and how all of this relates to your local Codex CLI workflows.

The Agent HQ Timeline

Agent HQ launched in public preview on 4 February 2026, bringing Claude and Codex alongside GitHub’s own Copilot cloud agent². At launch, each third-party agent ran on a single default model — there was no model picker. Over the following weeks, GitHub expanded the surface:

February 2026: Copilot Business and Pro users gained access to Claude and Codex agents³.
March 2026: GPT-5.3-Codex received long-term support status for GitHub Copilot, guaranteed available through February 2027⁴. GPT-5.4 became generally available for Copilot on 5 March⁵.
14 April 2026: Model selection shipped for Claude and Codex third-party agents, matching the existing Copilot model picker¹.

timeline
    title Agent HQ Evolution
    4 Feb 2026 : Agent HQ public preview
                : Claude + Codex agents (fixed models)
    26 Feb 2026 : Business/Pro access
    5 Mar 2026 : GPT-5.4 GA in Copilot
    18 Mar 2026 : GPT-5.3-Codex LTS
    14 Apr 2026 : Model selection for third-party agents

How Model Selection Works

When you open the Agents tab on an enabled repository, or assign a Codex task from an issue or PR, you now see a model picker dropdown. The available models for the Codex agent are¹:

Model	Strengths	Best For
GPT-5.4	Latest flagship, highest benchmark scores, strongest reasoning	Complex refactoring, architecture decisions, multi-file changes
GPT-5.3-Codex	Purpose-built for coding, 25% faster than GPT-5.2, LTS until Feb 2027	Bread-and-butter feature work, reliable daily driver
GPT-5.2-Codex	Legacy, being deprecated for ChatGPT sign-in	Only if you need reproducibility with existing workflows

There is also an Auto option that lets GitHub select the model based on task characteristics⁶.

For the Claude agent, the picker offers Claude Opus 4.6, Claude Sonnet 4.6, Claude Opus 4.5, and Claude Sonnet 4.5¹.

The Premium Request Cost Model

Every agent session on GitHub consumes premium requests from your Copilot subscription allowance⁷. The cost varies by model:

A standard chat interaction with the default model costs 1 premium request⁷.
More capable models like GPT-5.4 or Claude Opus 4.6 consume more premium requests per interaction⁷.
Complex multi-step agent tasks may consume several premium requests in a single session⁷.

This creates a direct cost-quality trade-off. Running every issue through GPT-5.4 burns through your monthly allowance faster than routing routine tasks to GPT-5.3-Codex.

GitHub Agent HQ vs Local Codex CLI

Understanding the architectural differences between GitHub’s hosted Codex agent and your local Codex CLI is essential for choosing the right surface for each task.

flowchart LR
    subgraph GitHub["GitHub Agent HQ"]
        I[Issue / PR / Agents Tab] --> CS[Cloud Sandbox]
        CS --> DR[Draft PR]
        DR --> R[Human Review]
    end
    subgraph Local["Local Codex CLI"]
        P[Terminal Prompt] --> LS[Local Sandbox]
        LS --> FC[File Changes]
        FC --> C[Commit / Push]
    end
    style GitHub fill:#e8f4f8,stroke:#0366d6
    style Local fill:#f0f8e8,stroke:#28a745

Dimension	GitHub Agent HQ	Local Codex CLI
Execution	Asynchronous, cloud sandbox	Interactive or `codex exec`, local sandbox
Internet	Disabled during execution⁸	Configurable (sandbox network policy)
Model choice	GPT-5.2/5.3/5.4 via picker	Any model via `config.toml` + custom providers
Context	Repository clone + setup script	Full filesystem, MCP servers, skills, subagents
Output	Draft PR for review	Direct file modifications
Multi-agent	Assign same issue to Copilot + Claude + Codex	TOML subagents, path-based addressing, up to 6 threads
Cost	Premium requests (Copilot subscription)	ChatGPT subscription or API key billing
Customisation	`.agent.md` files, AGENTS.md	config.toml, AGENTS.md, skills, hooks, plugins

When to Use Each

Use GitHub Agent HQ when:

You want fire-and-forget task execution on an existing issue
The task is self-contained within the repository (no external services needed)
You want to compare how Copilot, Claude, and Codex each approach the same problem
You are on mobile or away from your development machine

Use local Codex CLI when:

You need MCP server access (Jira, databases, monitoring tools)
The task requires interactive steering or mid-turn correction
You need subagent orchestration with more than one parallel thread
You want full control over sandbox policy, hooks, and approval modes

Custom Agents: The `.agent.md` Format

Alongside the third-party agent model picker, GitHub’s custom agent system lets you define specialised agents as .agent.md files in your repository⁹. These are not Codex CLI skills or plugins — they are GitHub Copilot cloud agents with their own instructions, tool access, and MCP server configurations.

File Structure

Custom agents live in .github/agents/ at the repository level, or in .github-private/agents/ for organisation-wide availability⁹:

.github/
  agents/
    security-reviewer.agent.md
    migration-helper.agent.md
    test-writer.agent.md

YAML Front Matter

Each .agent.md file combines YAML front matter with Markdown instructions⁹:

---
name: "security-reviewer"
description: "Reviews code changes for security vulnerabilities, OWASP Top 10, and dependency risks"
tools:
  - "github_api"
  - "code_search"
model: "gpt-5.4"
target: "github-copilot"
mcp-servers:
  snyk:
    command: "npx"
    args: ["-y", "@snyk/mcp-server"]
---

# Security Reviewer

You are a security-focused code reviewer. For every PR assigned to you:

1. Check for OWASP Top 10 vulnerabilities
2. Verify dependency versions against known CVEs
3. Flag any secrets or credentials in the diff
4. Provide a risk assessment (Low/Medium/High/Critical)

Key constraints: the prompt body has a maximum of 30,000 characters, and filenames may only contain periods, hyphens, underscores, and alphanumerics⁹.

How Custom Agents Relate to Codex CLI Skills

Custom .agent.md files and Codex CLI SKILL.md files solve similar problems — specialised agent behaviour for specific tasks — but they operate on different surfaces:

Aspect	`.agent.md` (GitHub)	`SKILL.md` (Codex CLI)
Runtime	GitHub Copilot cloud agent	Local Codex CLI session
Discovery	Agents tab picker	Automatic via directory walk
MCP	Configured per-agent in YAML	Configured in `config.toml`
Model	Set in front matter (`model:` key)	Inherited from session or profile
Invocation	Manual selection or `@mention`	Implicit (description match) or explicit
Max size	30,000 characters	No hard limit (32 KiB AGENTS.md ceiling applies separately)

For teams running both GitHub Agent HQ and local Codex CLI, the practical pattern is to maintain .agent.md files for cloud-executed workflows (CI review, issue triage) and SKILL.md files for interactive development tasks.

The Multi-Agent Bake-Off Pattern

One of Agent HQ’s most distinctive capabilities is assigning the same issue to multiple agents simultaneously². You can assign an issue to Copilot, Claude, and Codex, then compare the three draft PRs they produce.

This is architecturally different from Codex CLI’s multi-agent model. On GitHub, each agent runs independently with no inter-agent communication — they are parallel, isolated attempts at the same problem. In Codex CLI, subagents communicate via structured messaging, share context through the parent agent, and can be orchestrated as a coordinated team.

Practical Workflow

flowchart TD
    IS[GitHub Issue: 'Refactor auth module'] --> A1["Assign @Copilot (GPT-5.4)"]
    IS --> A2["Assign @Claude (Opus 4.6)"]
    IS --> A3["Assign @Codex (GPT-5.3-Codex)"]
    A1 --> PR1[Draft PR #1]
    A2 --> PR2[Draft PR #2]
    A3 --> PR3[Draft PR #3]
    PR1 --> REVIEW[Compare approaches]
    PR2 --> REVIEW
    PR3 --> REVIEW
    REVIEW --> MERGE[Merge best solution]

This pattern works well for:

Architecture decisions — see how different models approach the same structural problem
Evaluating model upgrades — compare GPT-5.3-Codex vs GPT-5.4 on your actual codebase
Critical changes — get independent verification from different AI providers

The cost is three sessions’ worth of premium requests per issue, so reserve this for high-value decisions rather than routine work.

Model Selection Strategy for GitHub Agent Tasks

With three Codex models now available on GitHub, a deliberate selection strategy prevents both wasted premium requests and suboptimal results.

Decision Framework

flowchart TD
    T[New GitHub Agent Task] --> C{Task complexity?}
    C -->|Simple bug fix, docs update| M1["GPT-5.3-Codex<br/>(reliable, lower cost)"]
    C -->|Multi-file refactor, architecture| M2["GPT-5.4<br/>(strongest reasoning)"]
    C -->|Reproducing existing behaviour| M3["GPT-5.2-Codex<br/>(legacy compatibility)"]
    C -->|Unsure| M4["Auto<br/>(GitHub selects)"]

GPT-5.4 is the right choice when the task requires reasoning across multiple files, understanding architectural constraints, or making non-obvious design decisions. It consumes more premium requests but produces better results on complex tasks⁵.

GPT-5.3-Codex remains the workhorse for standard feature implementation. Its LTS status until February 2027 means your team can standardise on it without worrying about imminent deprecation⁴. It is 25% faster than GPT-5.2-Codex and purpose-built for coding tasks.

GPT-5.2-Codex is deprecated for ChatGPT sign-in users as of 14 April 2026¹⁰, but remains available on GitHub. Use it only if you have existing workflows that depend on its specific behaviour and you need time to migrate.

Connecting GitHub Agent Work to Local Codex CLI

The most effective teams use both surfaces. A common pattern:

Triage on GitHub: assign routine issues to the Codex agent with GPT-5.3-Codex
Complex work locally: pull down the branch for architectural changes, use Codex CLI with MCP servers and subagents
Review on GitHub: use custom .agent.md reviewers or assign review to Claude for cross-model verification
CI validation: run codex exec in GitHub Actions via openai/codex-action for automated checks

The AGENTS.md file in your repository root serves both surfaces — GitHub’s Codex agent reads it when cloning the repo, and your local Codex CLI reads it during interactive sessions. This shared context file is the bridge between cloud and local workflows.

Enterprise Considerations

For Copilot Business and Enterprise subscribers, administrators must explicitly enable the Claude and Codex agent policies before team members can use them¹. This provides organisational control over which AI providers are permitted.

Enterprise teams should consider:

Model governance: standardise on GPT-5.3-Codex LTS for predictability, allow GPT-5.4 for senior engineers working on complex tasks
Premium request budgets: monitor consumption via the Copilot billing dashboard, set alerts at budget thresholds⁷
Custom agent libraries: build a shared .github-private/agents/ directory with organisation-specific review agents, security checkers, and compliance validators⁹
Audit trails: agent sessions produce PR comments and draft PRs, creating a natural audit log of AI-assisted work

What Is Still Missing

Despite the model selection addition, several gaps remain:

No subagent orchestration on GitHub: unlike Codex CLI’s TOML-defined subagents with path-based addressing, GitHub agents run as isolated sessions with no inter-agent coordination
No MCP server persistence: custom agents can define MCP servers, but the cloud sandbox’s internet restrictions limit what external services can be reached during execution⁸
No mid-session model switching: once a task starts with a selected model, you cannot change models mid-session (unlike Codex CLI’s /model command)
No Spark model: GPT-5.3-Codex-Spark (1,000+ tokens/second on Cerebras WSE-3) is not available on GitHub — it requires a ChatGPT Pro subscription and runs only through the local CLI or Codex App ⚠️

Looking Ahead

GitHub is actively working with Google, Cognition (Devin), and xAI to bring more agents into Agent HQ². The custom agent .agent.md format is evolving toward a richer configuration surface. The model picker is likely to expand as new models launch — when GPT-5.4-mini or future Codex-specific models arrive, they should appear in the picker automatically.

For Codex CLI practitioners, the key insight is that GitHub Agent HQ is not a replacement for local CLI workflows — it is a complementary surface optimised for asynchronous, issue-driven work within the GitHub ecosystem. The model selection feature closes a significant gap, giving you the same model routing flexibility on GitHub that you already have locally via config.toml profiles.

Citations

GitHub Blog, “Model selection for Claude and Codex agents on github.com,” GitHub Changelog, 14 April 2026. https://github.blog/changelog/2026-04-14-model-selection-for-claude-and-codex-agents-on-github-com/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶
GitHub Blog, “Pick your agent: Use Claude and Codex on Agent HQ,” 4 February 2026. https://github.blog/news-insights/company-news/pick-your-agent-use-claude-and-codex-on-agent-hq/ ↩ ↩² ↩³
GitHub Blog, “Claude and Codex now available for Copilot Business & Pro users,” GitHub Changelog, 26 February 2026. https://github.blog/changelog/2026-02-26-claude-and-codex-now-available-for-copilot-business-pro-users/ ↩
GitHub Blog, “GPT-5.3-Codex long-term support in GitHub Copilot,” GitHub Changelog, 18 March 2026. https://github.blog/changelog/2026-03-18-gpt-5-3-codex-long-term-support-in-github-copilot/ ↩ ↩²
GitHub Blog, “GPT-5.4 is generally available in GitHub Copilot,” GitHub Changelog, 5 March 2026. https://github.blog/changelog/2026-03-05-gpt-5-4-is-generally-available-in-github-copilot/ ↩ ↩²
GitHub Docs, “About third-party agents,” 2026. https://docs.github.com/en/copilot/concepts/agents/about-third-party-agents ↩
GitHub Docs, “Requests in GitHub Copilot,” 2026. https://docs.github.com/en/copilot/concepts/billing/copilot-requests ↩ ↩² ↩³ ↩⁴ ↩⁵
OpenAI, “Introducing Codex,” 2026. https://openai.com/index/introducing-codex/ ↩ ↩²
GitHub Docs, “Creating custom agents for Copilot cloud agent,” 2026. https://docs.github.com/en/copilot/how-tos/use-copilot-agents/cloud-agent/create-custom-agents ↩ ↩² ↩³ ↩⁴ ↩⁵
OpenAI Developers, “Changelog — Codex,” April 2026. https://developers.openai.com/codex/changelog ↩

GitHub Agent HQ Model Selection: Choosing GPT-5.4 vs GPT-5.3-Codex, Custom Agents, and the Multi-Agent GitHub Workflow

The Agent HQ Timeline

How Model Selection Works

The Premium Request Cost Model

GitHub Agent HQ vs Local Codex CLI

When to Use Each

Custom Agents: The .agent.md Format

File Structure

YAML Front Matter

How Custom Agents Relate to Codex CLI Skills

The Multi-Agent Bake-Off Pattern

Practical Workflow

Model Selection Strategy for GitHub Agent Tasks

Decision Framework

Connecting GitHub Agent Work to Local Codex CLI

Enterprise Considerations

What Is Still Missing

Looking Ahead

Citations

Custom Agents: The `.agent.md` Format