Chat Playground

The playground at /playground is where you talk to agents directly. On the surface it looks like a familiar chat interface, but the value is in the developer-oriented side panel: every message you send produces a full execution trace you can inspect, every agent setting can be temporarily overridden for the current session without modifying the saved agent, and the model picker is wired into the same registry that powers every other screen on the platform. The playground is the fastest path from "I changed something in my agent" to "I see exactly what that change did".

Using your own agents

Selecting an agent from the dropdown at the top of the playground loads its full saved configuration: provider, model, system prompt, knowledge bases, skills, tools, guardrails, and memory. From that point on, every message you send is processed exactly as it would be in production — same gateway, same tools, same guardrail decisions, same trace. This is the single most important property of the playground: it is not a sandbox with different behaviour from production; it is production with a chat UI in front of it.

• Multiple chat sessions are listed in the sidebar, each preserving its own conversation history and memory state. Switching between sessions is instant and every session is saved indefinitely so you can come back weeks later and see what you were working on.
• File and document attachments can be dropped directly into the chat. If the agent has a knowledge- base search tool, the platform automatically chunks and embeds the dropped file into a per-session ephemeral knowledge base; if the agent has a code interpreter, CSV files become directly queryable; otherwise the file content is added inline to the conversation context.
• Per-session overrides let you tweak the temperature, swap the model, edit the system prompt, or even disable an attached tool for the current session only. The overrides are clearly marked in the trace so you do not confuse experimental behaviour with saved behaviour, and the overrides disappear the moment you start a new session.
• Skill sample tour and template tour are short interactive overlays that walk you through how to use a particular skill or template inside the playground, with example prompts you can run with one click.

The built-in AgentSwarms Assistant

Available at the top of the agent picker on every playground page is a built-in agent we maintain and tune ourselves called the AgentSwarms Assistant. It is a friendly, always-available agent that knows two things really well: the field of agentic AI as a whole, and the AgentSwarms platform in particular. Using it is the fastest way to learn either, and we deliberately built it because a chat-based help interface is far more useful for a learning platform than yet another static documentation page.

• For conceptual questions like "explain ReAct versus Plan-and-Execute with a swarm I can run", the assistant gives the conceptual answer and offers to provision the relevant template into your workspace so you can see the pattern executing within seconds of reading about it. This loop of "explain then run" is the single most pedagogically effective interaction we ship.
• For platform questions like "how do I add a critic loop to my swarm", the assistant gives step-by-step instructions with deep links into the exact canvas controls you need to click. The instructions are kept up to date by attaching this documentation as a knowledge base to the assistant itself, so the assistant's answers always reflect the current state of the platform.
• For debugging questions like "why is my agent's tool not being called", the assistant asks you to share a trace ID, then opens the trace, reasons through what happened step by step, and explains the most likely root cause. This is the recommended escalation path before contacting support.
• For model selection questions like "what is the cheapest model that supports tool calling and a 128k context window", the assistant queries the live model registry, filters by your constraints, and returns a ranked list with current prices and a one-click "use this in my agent" action.

Why a friendly assistant matters for learning

The historic problem with learning a fast-moving technical field is that the gap between "I have a specific small question" and "I find a satisfying answer to that question" is usually measured in hours of Googling, reading conflicting Stack Overflow answers, and skimming blog posts of varying quality. A well-tuned in-platform assistant collapses that gap to seconds and removes a major source of attrition for learners. The AgentSwarms Assistant is not a marketing feature; it is the way we believe most people should learn agentic AI in 2026, and it is one of the most-used surfaces on the platform for that reason.