Multi Agent Orchestration Patterns

Sketchnote: Multi-Agent Orchestration Patterns for Codex CLI

Multi-Agent Orchestration Patterns for Codex CLI

Published: 2026-05-18 Source: Addy Osmani — “The Code Agent Orchestra” (https://addyosmani.com/blog/code-agent-orchestra/)

Summary

A synthesis of Addy Osmani’s analysis of what makes multi-agent coding work, mapped to Codex CLI capabilities.

Three Coordination Patterns

1. Subagents (Simple Delegation)

  • Parent decomposes work, spawns focused children with specific briefs
  • No peer messaging — parent collects all results
  • Codex CLI native: use custom agents in ~/.codex/agents/ or .codex/agents/
  • Cost: ~220k tokens for a three-agent task

2. Agent Teams (True Parallelism)

  • Shared task list with dependency tracking
  • Peer-to-peer messaging between teammates
  • Sweet spot: 3-5 agents; costs scale linearly
  • In Codex: combine subagents with git worktrees for isolation

3. Hierarchical Subagents

  • Feature leads spawn their own specialists
  • Mimics real org structure — prevents context fragmentation
  • Codex: set max_depth = 2 (carefully) to allow one level of sub-delegation

Four Essential Quality Gates

  1. Plan approval — agents write implementation plans before coding; review before code exists
  2. Loop guardrails — set max iterations with forced reflection on failure
  3. Dedicated reviewer agent — read-only tools, lint/test only, auto-triggered on task completion
  4. Lifecycle hooks — automated checks: tests must pass, linting before merge

Key Insight: The Bottleneck Shifted

“The bottleneck is no longer generation. It’s verification.”

Agents produce impressive output at speed, but human confidence in correctness is the hard constraint. Quality gates aren’t overhead — they’re the safety system.

Practical Recommendations

AGENTS.md Discipline

  • Human-curated only (LLM-generated versions show ~3% performance degradation)
  • Document: style, gotchas, architecture decisions, test strategies
  • Acts as institutional memory across sessions

Multi-Model Routing

  • Planning/architecture → cheaper models (GPT-5.4 mini)
  • Implementation → GPT-5-codex
  • Review → specialised reviewer agent with strict instructions

The Factory Model

  • View as a pipeline: Plan → Spawn → Monitor → Verify → Integrate → Retro
  • WIP limits: 3-5 concurrent agents
  • Kill criteria: stuck after 3+ iterations
  • Granular commits on feature branches

Mapping to Codex CLI

Osmani’s Pattern Codex CLI Implementation
Subagents Custom agent TOML files + spawn_agents_on_csv
Agent teams Multiple subagents + git worktrees
Hierarchical max_depth = 2 with feature-lead agents
Plan approval Architect agent reviews before worker agents start
Reviewer agent sandbox_mode = "read-only" custom agent
Lifecycle hooks /hooks configuration in config.toml
Multi-model routing Per-agent model override in TOML