agent▌

Create new agent skills with proper structure, progressive disclosure, and bundled resources.

Multi-agent architectures distribute work across multiple language model instances, each with its own context window. When designed well, this distribution enables capabilities beyond single-agent limits. When designed poorly, it introduces coordination overhead that negates benefits. The critical insight is that sub-agents exist primarily to isolate context, not to anthropomorphize role division.

Trust system reminders as agent progress notifications. Don't poll.

When the user asks to implement something, use implementation agents to preserve main context.

Multi-agent architectures distribute work across multiple agent invocations, each with its own focused context. When designed well, this distribution enables capabilities beyond single-agent limits. When designed poorly, it introduces coordination overhead that negates benefits. The critical insight is that sub-agents exist primarily to isolate context, not to anthropomorphize role division.

Playwright-based browser automation with element refs and session persistence for AI agents. \n \n Provides 6 core functions: open (navigate + configure), snapshot (refresh element refs), interact (click/fill/drag/upload/scroll), screenshot, execute (JavaScript), and close \n Uses @e ref system for element targeting; refs invalidate after navigation and require re-snapshot to refresh \n Supports video recording with optional cursor indicator, proxy routing, file uploads, and drag-and-drop intera

When working on the Agents workbench layout, always follow these guidelines:

Before dispatching, analyze the task systematically. Think through step by step:

Manage multiple local CLI agents in parallel tmux sessions with task assignment and monitoring. \n \n Start, stop, and monitor agents running in isolated tmux sessions with log tailing and health checks \n Assign tasks to specific agents and track their execution output in real time \n Schedule recurring agent work via cron integration for automated workflows \n Requires tmux and Python 3; agents are configured in a local agents/ directory \n

Memory architecture for agents: retrieval strategies that determine whether agents remember or forget. \n \n Covers five memory types: short-term (context window), long-term (vector stores), working memory, episodic memory, and semantic memory, each suited to different information patterns \n Emphasizes retrieval as the core challenge; provides chunking strategies, embedding quality guidance, and metadata filtering to surface the right memories at decision time \n Includes anti-patterns like sto