AgentSwarms is a browser-based platform for learning and prototyping agentic AI. You build multi-agent workflows on a visual node-graph canvas, wire up patterns like RAG, tool use, planner-executor, and human-in-the-loop, and watch runs execute live with full tracing. It combines a hands-on curriculum with a playground and 17 free developer tools. Best suited for students, engineers, and evaluators who want to learn how agents work by building them directly in the browser.

• Visual Swarm Canvas: Drag-and-drop node-graph builder to connect specialized agents into parallel or sequential workflows without writing backend code.
• Hands-On Curriculum: 50+ lessons, 97 notebooks, 21 interactive decks, and 25 build-along labs across 8 tracks from foundations to scaling.
• Seven Agentic Patterns: Live implementations of tool use, RAG, planner-executor, reflection, routing, parallel fan-out, and HITL approvals.
• Failure-Mode Labs: Diagnose and repair deliberately broken swarms with hallucinating RAG, runaway loops, and dead branches to build real debugging skills.
• Enterprise Templates: Runnable architectures for SOC triage, financial earnings analysis, and code review pipelines based on real production systems.
• MCP Support: First-class Model Context Protocol integration so you can connect any compliant tool or data source without bespoke wiring.
• Observability & Tracing: Every run logs prompts, model thinking, tool calls, tokens, latency, and per-node cost with clickable edge tracing.
• Voice Agents & BYOK: Add voice to any agent and bring your own keys for OpenAI, Anthropic, Gemini, Bedrock, Azure, OCI, Qwen, or Grok.

• Free During Beta: Full access to curriculum, playground, and included AI gateway credits with no credit card required.
• Learning-First Approach: Guided lessons explain concepts before you build, so the visual canvas actually makes sense to beginners.
• Real Production Patterns: Templates mirror architectures used inside Uber QueryGPT, Snowflake Cortex, and Salesforce Agentforce.
• No Setup Required: In-browser execution with local DuckDB WASM means you skip backend configuration entirely.
• BYOK Flexibility: Plug in your own provider keys to avoid daily gateway limits and pay providers directly at their rate.
• Deep Observability: Click any edge to inspect the JSON passing between agents, useful for understanding and debugging swarm behavior.

• Not a Production Runtime: Positioned as a learning and POC platform, not for fully unattended or production agent deployments.
• Daily Gateway Limits: Included AI credits reset daily, so heavy users need to add their own API keys to keep working.
• Focused on Learning Over Full Application Building: AgentSwarms focuses on teaching agent concepts and workflows. Users looking for a complete production AI application platform with advanced deployment, hosting, and team operations may need additional tools.
• Limited Team Features: Shared workspaces, admin controls, and centralized billing aren't available yet.

• Learning Agentic AI: Students and engineers work through 8 tracks covering foundations, memory, SQL agents, and multi-agent swarms with live runnable examples. You read a concept then build it immediately.
• Prototyping Multi-Agent Workflows: Engineers wire up specialized nodes for data extraction, sentiment analysis, or code review to test whether a swarm approach fits their problem before writing production code.
• Debugging Agent Failures: Developers use Failure-Mode Labs to practice fixing hallucinating RAG, runaway loops, and dead branches in a safe sandbox.
• Evaluating Enterprise Patterns: Technical evaluators run pre-built templates for SOC triage, earnings analysis, or code review pipelines to compare architectures against their own use case.
• Testing MCP Integrations: Developers connect Model Context Protocol servers to see how tools and data flow into agent workflows without building custom integrations.
• Building Voice Agents: Teams prototype speech-to-speech agents like support triage or B2B discovery callers using the Voice Playground.