Multi-agent coding patterns: from single AI to orchestrated teams

Addy Osmani's post is a written companion to a talk he delivered at O'Reilly AI CodeCon, covering the landscape of patterns for coordinating AI coding agents in real-world software workflows. He frames the central shift as moving from the "conductor" model — one agent, synchronous, constrained by a single context window, using tools like Claude Code CLI and Cursor in-editor agent mode — to the "orchestrator" model, where multiple agents work asynchronously, each with its own context window and file scope, using tools like Agent Teams, Conductor, Codex, and Copilot Coding Agent. The developer's role changes from writing code in real time to planning work, decomposing tasks, assigning them, and verifying outputs.

Osmani identifies three hard ceilings that single-agent workflows inevitably hit: context overload (large codebases overwhelm one context window), lack of specialization (a generalist agent handling data, API, UI, and tests produces lower-quality output than a focused one), and lack of coordination (agents spawned without coordination primitives become harder to manage as their number grows). Multi-agent architectures address all three through four compounding advantages: parallelism (three agents building frontend, backend, and tests simultaneously), specialization (each agent scoped only to the files it owns, e.g., an agent that only knows `db.js` writes better database code), isolation via Git worktrees (each agent gets its own working directory, eliminating merge conflicts), and compound learning through an `AGENTS.md` file that accumulates patterns and gotchas across sessions. He references Steve Yegge's eight-level framework for developer AI maturity, noting that most developers are stuck at levels 3–4 and that the orchestration tier begins at level 6, requiring a fundamentally different skill set.

Key facts

1. 01Addy Osmani delivered this material as a talk at O'Reilly AI CodeCon, and the post is the written companion.
2. 02The 'conductor' model (single agent, synchronous) uses tools like Claude Code CLI and Cursor in-editor agent mode.
3. 03The 'orchestrator' model (multiple async agents) uses tools like Agent Teams, Conductor, Codex, and Copilot Coding Agent.
4. 04Steve Yegge's eight-level AI developer framework is used as a scaffold; the talk covers levels 5 through 8.
5. 05Most developers are described as stuck at levels 3–4; the orchestration tier starts at level 6.
6. 06Four multi-agent advantages are identified: parallelism, specialization, isolation (via Git worktrees), and compound learning via an `AGENTS.md` file.