agents▌

Maintain small, high-signal AGENTS.md files:

When launching parallel background agents, do NOT poll with sleep loops.

Orchestration through Claude Code's built-in Agent Tool

When launching multiple agents in parallel, follow this pattern to avoid context bloat.

When you have multiple unrelated failures (different test files, different subsystems, different bugs), investigating them sequentially wastes time. Each investigation is independent and can happen in parallel.

The fundamental architecture of computer use agents: observe screen, reason about next action, execute action, repeat. This loop integrates vision models with action execution through an iterative pipeline.

Build stateful AI agents on Cloudflare Workers with WebSockets, persistent state, scheduling, and multi-agent coordination. \n \n WebSocket-based real-time communication with automatic state synchronization across clients and devices; resumable streaming persists across disconnects and page refreshes \n Built-in SQLite storage (up to 1GB per agent), task scheduling with cron expressions, and Durable Objects for globally unique, persistent agent instances \n Multi-agent coordination via routeAgen

When you have multiple unrelated failures (different test files, different subsystems, different bugs), investigating them sequentially wastes time. Each investigation is independent and can happen in parallel.

Build text and voice agents with tools, multi-agent handoffs, guardrails, and human-in-the-loop patterns. \n \n Supports text agents, realtime voice agents with WebRTC, and multi-agent workflows with automatic delegation via handoffs \n Define tools using Zod schemas for type-safe parameter validation; includes structured output support for predictable JSON responses \n Implements input/output guardrails for safety validation, human approval workflows for sensitive operations, and streaming supp

Natural conversation with AI through speech, balancing latency against control. \n \n Choose between speech-to-speech models (lowest latency, less controllable) or pipeline architectures (STT→LLM→TTS for fine-grained control) \n Core challenges: latency budgeting across all components, voice activity detection, barge-in handling, and turn-taking to avoid awkward pauses or overlaps \n Requires semantic VAD, response length constraints in prompts, and noise handling to achieve natural conversation