Code Review

AgentSwarms — Our Biggest Release Yet: 100+ Executable Framework Notebooks, Voice Agents, Website Embeds & More TL;DR — what's new: -103 executable, in-browser notebooks across 9 agent frameworks — no setup, run every cell live. -Voice agents you build with speech-to-text + text-to-speech and talk to right in the browser. -Website embeds — drop a standalone agent or a multi-agent task bot onto your own site with one iframe. -30 interactive field-engineering blogs + 25 interactive presentation decks -9 hands-on failure-mode labs + a 38-cell failure-modes notebook covering 12 real production failures -17 free, no-login interactive AI tools -Detailed swarm execution traces — tokens, cost, latency, every handoff -20 runnable industry templates AgentSwarms is a learn-by-building platform for Agentic AI: build, run, test, and ship AI agents and multi-agent swarms — and actually learn the frameworks by running real code, not reading about it. This is our largest update to date. Here's everything that's new. 1. Framework-specific notebooks with executable cells — learn every major agent framework by running it We shipped 103 notebooks with live, editable cells that execute right in your browser — zero installs, zero API keys to wrangle. There are dedicated, in-depth tracks for OpenAI Agents SDK, Google ADK, LlamaIndex, Mastra, LangChain, LangGraph.js, and AWS Strands SDK — plus Vercel AI SDK and VoltAgent. Each track walks the whole arc: fundamentals → typed tools → multi-agent handoffs → memory → guardrails → evaluation, with clear, beginner-friendly explanations next to code you can change and re-run. Start here: https://agentswarms.fyi/notebooks 2. Build voice agents with TTS + STT — give any agent a voice and talk to it You can now turn any agent into a voice agent using a speech-to-text → LLM → text-to-speech pipeline, pick from 11 studio voices, and talk to it live in the Voice Playground — tap the mic, speak, and hear it answer back. Fork a ready-made sample to get started in seconds. Because voice runs on the same engine as chat, your tools, RAG, memory, and guardrails all carry straight over. Try it: https://agentswarms.fyi/voice-playground 3. Build an AI backend and embed agents + multi-agent task bots on your website Ship what you build. Embeds turns any agent or swarm into a widget you paste into your own site with a single iframe — no backend to stand up. Choose an Agent embed (a streaming chat panel that answers with the agent's full brain) or a Swarm embed (a run-and-return task bot where a whole multi-agent workflow executes server-side and returns the result). Every embed gets domain-locked API keys, a per-embed daily request cap, and serves anonymous visitors with no login. Details: https://agentswarms.fyi/docs/embeds 4. New interactive field-engineering blogs and presentations Our library now includes 30 field-engineering blog posts, each with custom interactive visuals you can click through — covering AI agents in financial services, multi-agent simulation, deploying MCP servers to AWS/Azure/GCP, voice agents, security, retrieval, cost control, and framework deep-dives. Alongside them are 25 interactive presentation decks on everything from LLM internals to multi-agent orchestration and LLMOps. Read them: https://agentswarms.fyi/blog 5. A lot more failure-mode labs and notebooks Real agents fail in ways demos never show, so we built the practice ground for it: 9 hands-on failure-mode labs in the swarm canvas that inject a real fault — a coordination deadlock, a cascading error, a runaway loop, a prompt injection — and then grade your fix. They're backed by a 38-cell "Failure Modes Lab" notebook that reproduces 12 production failure modes (hallucinations, sycophancy, RAG poisoning, context rot, cost blow-ups, goal drift, and more) and ships the exact fix for each. Start at https://agentswarms.fyi/learn 6. Free interactive AI tools We added 17 free, no-login browser tools for everyday agent work: a RAG chunking visualizer, prompt-injection tester, semantic chunker, swarm-architecture visualizer, PII redactor, multi-agent token-cost calculator, LLM-as-a-judge builder, tool-schema generator, GraphRAG triplet extractor, synthetic RAG-eval dataset generator, and more. Open and use them instantly: https://agentswarms.fyi/tools 7. Detailed swarm execution traces Every multi-agent run is now fully traced and replayable. Step through each node's inputs and outputs, every tool call and handoff, and read per-run token usage, cost, and latency on a timeline — the raw material for debugging, evals, and cost dashboards. No more guessing why a swarm did what it did. 8. More runnable industry templates We expanded the template library to 20 one-click, runnable industry templates — support triage / knowledge-base agent, research RAG, sales outreach, code reviewer, CRM data cleanser, long-context analysis, and more. Each spins up a working agent or swarm you can run immediately and adapt. Browse them: https://agentswarms.fyi/templates