SUMM

Sub-agents were introduced to Cloud Code to optimize token usage and improve code performance by offloading research tasks.
Before sub-agents, the main agent handled all tasks, leading to large context windows and frequent summarization, causing performance drops.
Sub-agents can perform tasks like scanning codebases and summarizing findings, reducing the number of tokens by only passing back essential summaries.
The system keeps sub-agent actions out of the parent’s conversation history, saving space and context for the main task.

Problems arise when sub-agents are tasked with implementation rather than just research or planning.
If sub-agents (e.g., front-end and back-end agents) implement separately, they lack shared context, making bug fixes and coordination difficult.
Each sub-agent operates in an isolated task session, leading to a loss of history and inability to know what other agents or sessions have done.
Currently, context sharing between agents is limited, but future improvements are expected.

Use sub-agents mainly as researchers to gather information and return concise summaries, not for direct implementation.
Expert sub-agents can be created per service (e.g., Vercel SDK, Stripe, Tailwind) and loaded with specialized documentation and tools.
Adam Wolf from the Cloud Code team supports the idea that sub-agents work best when delivering small, high-value research summaries.

Inspired by techniques from the Manus team, local files (e.g., Markdown files) are used to store task results and research, ensuring persistent, accessible context.
The parent agent maintains a context file (e.g., doc/task/) summarizing the project status and steps done.
All sub-agents read the current context file before starting work and update it with new findings or plans after task completion.
Research reports and implementation plans are saved to doc files so other agents or parent agents can consult them as needed.

Important documentation and references are added directly to the sub-agent’s system prompt to ensure up-to-date compliance with best practices.
Sub-agents have access to specialized tools (e.g., MCP tools for information retrieval) specific to their domain.
Agents are configured by editing settings files, setting up specialized MCP servers, and defining roles and access.
The system prompt includes specific rules: agents propose detailed implementation plans, never implement directly, and always update/read context files accordingly.
Standardized output formats ensure clarity and process consistency across all sub-agents.

Demonstrates creating a Next.js project using specialized sub-agents (e.g., Chassian, Vercel AI SDK).
Parent agent manages project context and delegates tasks, ensuring agents read and update context files before and after work.
Each sub-agent performs research, retrieves examples and documentation, and compiles detailed plans in shared files.
Parent agent reads these plans and handles the actual implementation, retaining full project context for consistent results.
The process results in a high-fidelity, high-functioning application quickly assembled by combining agent research and central coordination.

Cloud code now supports background sessions, allowing monitoring and task persistence.
The demonstrated approach produces fully functional, high-quality UIs and integrations (e.g., connecting with Vercel SDK) with smooth, detailed user experiences.
The iterative, file-based approach improves success rates, quality, and maintainability when using AI agents for code tasks.
The presenter offers additional resources, club sessions, and templates for advanced sub-agent workflows.
Viewers are encouraged to join the AI builder club and access shared templates, hooks, commands, and best-practice discussions.

I was using sub-agents wrong... Here is my way after 20+ hrs test