AgentSwarms
Learn AI agents by building them.
Hands-on lessons, runnable notebooks and real multi-agent experiments. Build with the latest GPT, Claude and Gemini models, add Jev decisions, or talk to a GPT-Live voice agent.
Free during beta · 5 free AgentSwarms AI Gateway credits per day · No credit card
Learn here, in your browser
agentswarms.fyi is your learning lab: concepts, best practices, guided experiments and certification.
Build real use cases on your infrastructure
Explore the source-available, self-hostable unified Agentic AI & Business Intelligence platform.
Q3_Earnings_Call.pdf
Delegating tasks…
- Revenue:
- $12.4B
- EPS:
- $1.12
Analyzing CEO Q&A tone…
Final Memo Generated.
New ways to build and experiment
Release notesJev: decisions, not guesswork
Classify tickets, score leads and flag content with typed outputs. Pair Jev with a writing agent in a swarm.
Try Jev templatesGPT-Live 1: speak naturally
Real-time voice conversations with captions and interruptions. Start with the concierge or interview coach.
Try a voice agentMore models, one workspace
Experiment with the latest GPT, Claude and Gemini options. Mix models across your agent and swarm steps.
Build an agentTemplates
Runnable templates from real industry use cases
Start from a working multi-agent system, not a blank canvas. Each template ships wired agents, tools, memory, and a swarm you can run in one click — export in your choice of framework, modify and run in your choice of environment.
Operations Swarm
An orchestrator routes work across specialist agents — intake, research, and synthesis — handing off context at each step until the task is done.
Market Report Agents
Parallel agents pull signals, score sentiment, and a synthesis agent writes the final memo — the analyst workflow as a runnable swarm.
Ship-Code Agents
A release pipeline of agents plans the change, writes and reviews the diff, then gates the deploy on a human approval node before it ships.
Voice
Build agents you can actually talk to
Choose GPT-Live 1 for real-time conversation with interruptions, or Classic voice for a transcribe → agent → speech pipeline with tools and retrieval.
You speak
Mic captured in the browser
Speech-to-text
Transcribed by gpt-4o-transcribe
The agent thinks
LLM + tools + knowledge + guardrails
The agent speaks
Spoken back in a natural voice
Choose GPT-Live or Classic voice
Build real-time GPT-Live 1 agents with a voice and instructions, or customize the speech models in the Classic pipeline.
Match the engine to the use case
Classic voice uses the chat agent tool and retrieval pipeline. GPT-Live uses a separate real-time engine; it does not automatically inherit every Classic tool or guardrail.
Ready-to-talk samples
Try the GPT-Live hotel concierge or interview coach, or explore the Classic support and language-tutor samples.
Embeds
Drop your agents & swarms into any website
Put an agent or swarm on your website with an iframe. Built-in Claude and Jev models are supported; visitors use your account's allowance. The basic embed path uses saved prompts, not every Playground tool or guardrail.
Support Agent
AI Assistant
One snippet, agent or swarm
Paste an iframe on your site. Agent embeds return the saved model's response, including Claude text and Jev decisions. Swarm embeds execute their graph server-side and return the final output.
Domain-locked API keys
Each embed has its own sk_embed_ key, allowed-domains list, and daily request cap. Rotate or revoke in one click when you need to.
Anonymous visitors, your account
Nobody has to sign in. Sessions persist per browser so returning visitors resume. Model calls bill to your AgentSwarms usage — capped by your daily limit.
Why AgentSwarms
Most courses make you watch. We make you build and break.
This is a learning and POC platform, not a production runtime. The fastest way to actually understand agentic AI is to run it, wire it, and watch it fail.
A real builder, not a toy
The same visual Agent Builder and drag-and-drop swarm canvas you'd prototype with — guardrails, tools, memory, RAG, evals, and human approval included.
Learn by building, not watching
Every concept maps to a runnable agent or swarm you open in the browser. Read it, run it, fork it — no setup, no API keys to start.
We teach the failure modes
Diagnose and repair deliberately-broken swarms in Failure-Mode Labs — hallucinating RAG, runaway loops, dead branches. Almost nobody else teaches this.
Curriculum
From "what's an agent?" to "I shipped a swarm."
Every lesson is interactive — read a concept, then run a live agent that demonstrates it.
New ways to learn
More than a curriculum
Pick the format that fits how you learn — animated explainers, guided builds, or hands-on debugging.
Interactive Presentations
Animated, click-through decks on LLMs, prompting, RAG, agentic patterns, orchestration, security, observability, and inference internals.
Open presentationsDeck 04 · Agentic patterns
ReAct: reason, act, observe, repeat
Build-Along Labs
Step-by-step guides that build 10 standalone agents and 15 multi-agent swarms in the real builder — with diagrams, exact settings, and deep links.
Start building
Failure-Mode Labs
Fix swarms that are deliberately broken. The platform verifies your repair — the fastest way to build real intuition for what goes wrong.
Diagnose a swarmmax_iterations exceeded — agent re-called the same tool 3 times
Your job: find the missing stop condition and repair the loop.
Interactive Notebooks
Runnable TypeScript notebooks in your browser — foundations, advanced topics, and every major framework (LangChain, LlamaIndex, OpenAI Agents SDK, Vercel AI SDK, Google ADK), plus 10 reproduced failure-mode examples.
Open notebooksfnd-prompt-engineering.ipynb
1 · Anatomy of a chat request
Interactive Blogs
Long-form field-engineering posts with live widgets baked in — VRAM estimators, cost calculators, latency simulators. Drag the sliders and watch the answer change in real time.
Read the blogHow it works
Four steps from zero to swarm
No installs. No API keys to start. Open a demo, follow the guided prompts, then make it your own.
Try a Live Demo
Pick any template — Product Support, Research Assistant, Code Reviewer — and a working agent is provisioned in seconds.
Follow the Guided Tour
Each demo opens with a side-panel lesson. Suggested prompts walk you through RAG, guardrails, and approvals.
Fork & Experiment
Tweak the system prompt, swap models, wire up your own knowledge base. Break things — that's how you learn.
Build Your Own
Compose your own agents, chain them into a swarm, and watch traces light up in the observability dashboard.