Exploring Subagents in Claude Code with Addy Osmani

Why It Matters

Subagents can speed AI‑generated software by parallelizing tasks while preserving context, but manual orchestration limits scalability, directly impacting developer efficiency and project costs.

Key Takeaways

• Subagents enable parallel execution of independent code components.
• Parent orchestrator manually manages dependency graph among subagents.
• Context isolation per subagent reduces prompt token overload.
• Specialization allows focused generation of data layer and business logic.
• Lack of automatic messaging requires extra coordination effort.

Summary

In this walkthrough, Addy Osmani demonstrates Claude Code’s sub‑agent feature by prompting the model to build a simple bookmarks manager called Link Shelf using Express and SQLite. The parent orchestrator reads the single request, decomposes it into three sub‑agent briefs, and launches two agents in parallel to generate the data‑layer and business‑logic code.

The demo highlights how the orchestrator can run independent agents simultaneously, then spawn a third agent to create API routes once the first two have written their files. This parallelism reduces overall generation time, while each sub‑agent enjoys context isolation, preventing token overload and allowing specialization. However, the parent must manually track dependencies and sequence execution, as there is no built‑in messaging or automatic unblocking between agents.

Osmani notes, “We get parallel execution where it’s possible, but the parent is manually managing our dependency graph.” The example shows the parent reading existing files to understand interfaces before wiring everything together, illustrating both the power and the current friction of the approach.

The implication is clear: sub‑agents can accelerate AI‑driven code creation and keep prompts concise, but developers still need to handle orchestration logic. Until automated messaging and dependency resolution are added, large‑scale or highly interdependent projects may require significant manual coordination, affecting productivity and cost efficiency.

Original Description

Multi-agent orchestration in Claude Code is genuinely powerful, but it's not magic yet, demonstrates Google’s Addy Osmani in this clip from AI Codecon. Feed Claude Code a single prompt and a parent orchestrator will decompose it into parallel subagents, each handling a discrete piece of the project independently, before spinning up a third to wire everything together. Context isolation, specialization, parallel execution: It all works. But as Addy puts it, "We get parallelism where it's possible, but the parent is manually managing our dependency graph" with no shared task list, no peer-to-peer messaging between agents, and no automatic unblocking. Check out Addy's breakdown of where agentic coding delivers today and where it still has room to grow.

Follow O'Reilly on:

LinkedIn: https://www.linkedin.com/company/oreilly/

Facebook: http://facebook.com/OReilly

Instagram: https://www.instagram.com/oreillymedia

BlueSky: https://bsky.app/profile/oreilly.bsky.social