Codex CLI vs Claude Code Multi-Agent: Subagents, Agent Teams and the Protocol Gap

codex-vs-claudesubagentsagent-teamsprotocol-gapbenchmarkstoken-efficiencya2amcpconvergence

Codex CLI vs Claude Code Multi-Agent: Subagents, Agent Teams and the Protocol Gap

The two dominant terminal-native coding agents — OpenAI’s Codex CLI and Anthropic’s Claude Code — have each shipped multi-agent capabilities, but with fundamentally different architectural philosophies. Codex CLI offers TOML-defined subagents with explicit spawning, path-based addressing and batch processing. Claude Code counters with Agent Teams: peer-to-peer mailbox communication, shared task lists, and autonomous delegation. This article dissects both systems head-to-head, compares their protocol stacks, examines the benchmarks, and asks whether they are converging.

Architectural Philosophies

The split reflects each vendor’s trust model. Codex CLI prioritises explicit user control and auditability — you define agents in TOML, set hard concurrency limits, and spawn them deliberately1. Claude Code emphasises autonomous agent judgement: Agent Teams members claim work from a shared task list and communicate directly without routing through a parent2.

graph TB
    subgraph Codex CLI
        P[Parent Agent] -->|spawn| S1[Subagent A<br/>/root/agent_a]
        P -->|spawn| S2[Subagent B<br/>/root/agent_b]
        S1 -->|result| P
        S2 -->|result| P
    end
    subgraph Claude Code
        TL[Team Lead] -->|TeamCreate| T1[Teammate 1]
        TL -->|TeamCreate| T2[Teammate 2]
        T1 <-->|SendMessage| T2
        T1 -->|SharedTaskList| TL
        T2 -->|SharedTaskList| TL
    end

The diagram captures the core difference: Codex subagents report results back to a parent in a hub-and-spoke pattern; Claude Code teammates talk to each other in a mesh.

Codex CLI Subagents: Explicit Control

Codex subagents became generally available in March 2026, graduating from a feature-flag preview to a stable default3. Custom agents are defined as standalone TOML files placed in ~/.codex/agents/ (personal scope) or .codex/agents/ (project scope)1.

Configuration

Each TOML file requires three fields — name, description, and developer_instructions — with optional overrides for model, sandbox_mode, and mcp_servers that inherit from the parent session when omitted1.

# .codex/agents/security-reviewer.toml
name = "security-reviewer"
description = "Reviews code changes for security vulnerabilities"
developer_instructions = """
Analyse diffs for OWASP Top 10 vulnerabilities.
Report findings as structured JSON with severity levels.
"""
model = "o4-mini"
sandbox_mode = "locked-network"

Runtime Controls

Global orchestration settings live in the [agents] section of config.toml1:

Setting Default Purpose
max_threads 6 Concurrent open agent thread cap
max_depth 1 Nesting depth (prevents recursive delegation)
job_max_runtime_seconds 1800 Per-worker timeout for batch jobs

Built-in Roles and Batch Processing

Codex ships three built-in agent roles: default (general-purpose), worker (execution-focused), and explorer (read-heavy codebase navigation)1. For parallel workloads, the experimental spawn_agents_on_csv tool accepts a CSV file and an instruction template with {column_name} placeholders, spawning one worker per row1. Each worker must call report_agent_job_result exactly once.

Path-Based Addressing

Since the March 2026 multi-agent v2 release, subagents receive path-based addresses (e.g., /root/agent_a) and accept structured steering instructions mid-execution4. This is a significant step toward richer inter-agent communication, though it remains parent-to-child rather than peer-to-peer.

Claude Code Agent Teams: Autonomous Collaboration

Agent Teams shipped with Claude Opus 4.6 on 5 February 2026 as an experimental feature, enabled via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=12. Teams support 2–16 agents working on a shared codebase.

Architecture

The system comprises four components5:

  1. Team Lead — your main Claude Code session, which analyses tasks, creates teams via TeamCreate, and orchestrates the workflow.
  2. Teammates — independent Claude Code processes, each with its own full context window and tool access.
  3. Shared Task List — stored at ~/.claude/tasks/{team-name}/, providing coordination with statuses, ownership, and dependency tracking.
  4. Mailbox — the SendMessage tool enables direct peer-to-peer messaging between any teammates, or broadcasts to the entire team.

Agent Definition

Claude Code uses Markdown with YAML frontmatter in .claude/agents/, a deliberate contrast to Codex’s TOML6:

---
name: security-reviewer
description: Reviews code for security vulnerabilities
model: opus
---

# Security Reviewer

Analyse diffs for OWASP Top 10 vulnerabilities.
Report findings with severity levels.

When Subagents vs Teams

Claude Code maintains both patterns. Subagents (the Agent tool) are quick, focused workers that report back to the parent — similar to Codex’s model. Agent Teams are for tasks where teammates need to share findings, challenge each other, and coordinate autonomously2. The trade-off: Agent Teams consume 4–7× more tokens than single-agent sessions7.

Head-to-Head Comparison

Benchmarks

Benchmark Codex CLI Claude Code Notes
SWE-bench Verified ~80% (GPT-5.3-Codex) 80.9% (Opus 4.5) Essentially tied89
Terminal-Bench 2.0 77.3% 65.4% 12-point Codex advantage7
Blind Code Quality 25% win rate 67% win rate Claude produces cleaner code7

The benchmarks tell a split story. Raw coding ability on SWE-bench Verified is a dead heat. Terminal-native tasks (scripting, system admin, DevOps) favour Codex CLI decisively. But when human developers blindly evaluate code quality, Claude Code wins two-thirds of the time7.

Token Efficiency and Cost

OpenAI claims a 4× token efficiency advantage for Codex CLI7. In practice, a complex refactor consuming 1.5M tokens on Codex CLI required 6.2M tokens for a comparable task on Claude Code7. Combined with lower per-token pricing (GPT-5.3-Codex at $1.25/$10 per MTok vs Opus 4.6 at $5/$25 per MTok), Codex works out roughly 10× cheaper per equivalent coding task on API pricing7.

Context Window

Claude Code holds a significant advantage here: Opus 4.6 supports up to 1M tokens of context, versus Codex CLI’s 192K7. For large-scale refactoring across many files, that 5× context advantage matters.

graph LR
    subgraph Token Economics
        A[Codex CLI<br/>1.5M tokens<br/>~$4.50] --> C{Same Task}
        B[Claude Code<br/>6.2M tokens<br/>~$46.50] --> C
    end
    subgraph Context Capacity
        D[Codex CLI<br/>192K window] --- E[Claude Code<br/>1M window]
    end

The Protocol Gap: MCP and A2A

Both tools support MCP (Model Context Protocol), but with different depths.

MCP Support

Codex CLI supports MCP servers as both STDIO processes and streamable HTTP endpoints, configured via config.toml and managed through codex mcp add10. Claude Code has deeper MCP integration with over 3,000 tool integrations and 14+ lifecycle hook trigger points711.

The A2A Question

A2A (Agent-to-Agent), created by Google in April 2025 and now stewarded by the Linux Foundation’s Agentic AI Foundation, standardises how agents discover and communicate with each other across vendor boundaries12. Over 100 enterprises have adopted it13.

Neither Codex CLI nor Claude Code natively implements A2A as of April 2026. Both can access A2A endpoints through MCP bridge servers — community-built adapters that expose A2A agent communication as MCP tools14. But native A2A support would allow Codex subagents and Claude Code teammates to collaborate across vendor boundaries without bridging overhead.

graph TB
    subgraph "Current State"
        CX[Codex CLI] -->|native| MCP1[MCP Server]
        CC[Claude Code] -->|native| MCP2[MCP Server]
        CX -.->|via bridge| A2A1[A2A Endpoint]
        CC -.->|via bridge| A2A2[A2A Endpoint]
    end
    subgraph "Convergence Target"
        CX2[Codex CLI] -->|native| MCP3[MCP]
        CX2[Codex CLI] -->|native| A2A3[A2A]
        CC2[Claude Code] -->|native| MCP4[MCP]
        CC2[Claude Code] -->|native| A2A4[A2A]
        A2A3 <-->|interop| A2A4
    end

The Convergence Thesis

Despite different starting points, both tools are converging architecturally:

  1. Codex is gaining richer communication. Path-based addressing and structured messaging (March 2026) move Codex subagents from pure hub-and-spoke toward something resembling peer awareness4.
  2. Claude Code is gaining explicit controls. Agent Teams still exposes configuration for team sizes, task dependencies, and ownership — structured orchestration layered on top of autonomous behaviour5.
  3. Both need A2A. The multi-vendor world demands that agents from different providers coordinate. MCP solves tool access; A2A solves agent coordination. The first tool to ship native A2A gains a significant interoperability advantage12.

The deeper pattern: Codex users should learn from Claude’s mailbox and shared task list patterns for richer inter-agent coordination. Claude Code users should learn from Codex’s explicit concurrency controls and token efficiency for cost management at scale.

Decision Framework

Factor Choose Codex CLI Choose Claude Code
Token budget matters ✅ 4× more efficient  
Terminal/DevOps tasks ✅ 77.3% Terminal-Bench  
Code quality priority   ✅ 67% blind win rate
Large codebase context   ✅ 1M token window
Peer-to-peer agent work   ✅ Agent Teams mailbox
Batch parallel processing ✅ spawn_agents_on_csv  
MCP ecosystem breadth   ✅ 3,000+ integrations
Autonomous execution ✅ Kernel-level sandbox  

What Codex Users Should Watch

  • Agent Teams patterns are coming. Codex’s structured messaging is a stepping stone toward peer-to-peer. Expect a mailbox-like primitive in a future release.
  • A2A native support. When this lands, Codex subagents become first-class citizens in cross-vendor workflows.
  • The context window gap. At 192K vs 1M, Codex needs to close this for complex multi-file refactoring scenarios.

The multi-agent landscape is young — both architectures are evolving fast. The winning strategy is not to pick one tool permanently, but to understand both models deeply enough to use each where it excels.

Citations

  1. [Subagents – Codex OpenAI Developers](https://developers.openai.com/codex/subagents)

    ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  2. Claude Code Agent Teams: Setup & Usage Guide 2026 ↩ ↩2 ↩3

  3. Use subagents and custom agents in Codex — Simon Willison ↩

  4. Codex vs Claude Code: The Divergence in Subagent Design Philosophy — SmartScope ↩ ↩2

  5. From Tasks to Swarms: Agent Teams in Claude Code — alexop.dev ↩ ↩2

  6. Claude Code Custom Subagents: Complete Guide — Claude Lab ↩

  7. Codex vs Claude Code: Which CLI Agent Wins for Your Workflow in 2026 — Particula ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9

  8. SWE-Bench Verified Leaderboard March 2026 — Marco Patzelt ↩

  9. Best LLM for Coding 2026 — SmartScope ↩

  10. [Model Context Protocol – Codex OpenAI Developers](https://developers.openai.com/codex/mcp)

    ↩

  11. AI Weekly: Claude Code Dominates, MCP Goes Mainstream — DEV Community ↩

  12. MCP vs A2A: The Complete Guide to AI Agent Protocols in 2026 — DEV Community ↩ ↩2

  13. Deploying multi-agent systems using MCP and A2A with Claude on Vertex AI — Anthropic ↩

  14. A2A-MCP-Server — GitHub ↩