Agent Builder

New agents default to the built-in AgentSwarms AI provider, so you can save and chat with a working agent before configuring anything. This page walks the configuration blocks in the order they appear in the form.

Identity

Name and description follow the agent everywhere it is referenced — swarm nodes, playground pickers, community listings, the trace viewer. A descriptive name ("Insurance Claim Triage v2") pays for itself the first time you have more than a handful of agents.

Model

Provider

AgentSwarms AI (built-in, no key), OpenAI, Google Gemini, Grok (xAI), Groq, OpenRouter, Anthropic, AWS Bedrock, Google Vertex AI, Azure OpenAI, OCI Generative AI, Qwen (DashScope), custom Ollama, or self-hosted vLLM. Bring-your-own-key providers are configured in Integrations.

Model

Filtered by provider. The Model Registry supplies context-window and pricing data per model.

Temperature

Zero for deterministic tasks (classification, SQL); higher for creative drafting.

Sampling & limits

Max tokens, top-p, frequency penalty, presence penalty, and stop sequences — standard defaults pre-filled, ignorable unless you have an opinion.

System prompt

The single most important field on the form. Starter prompts can be pulled from the Prompt Library, and the same prompt patterns carry over when the agent is dropped into a swarm node.

Knowledge bases

Attach one or more knowledge bases from /knowledge and the agent automatically gains a search tool scoped to them — no separate tool setup. Knowledge bases are pgvector indexes populated from:

• File uploads — PDF, CSV, Markdown, plain text; parsed, chunked, and embedded automatically with live status.
• URL ingestion — pull a page's content in as a document, re-syncable from the sources panel.
• GitHub ingestion — index files from a public repo, also re-syncable.
• Pasted text — the fastest path for short policy docs and FAQs.

The knowledge page also has a Graph tab that builds an entity-relation graph over a knowledge base; agents can then use Knowledge Graph Search for multi-hop Graph-RAG questions.

Tools

Web Search

Live search with built-in Firecrawl, or bring your own key (Brave, SerpAPI, Tavily).

Web Browser

Fetches a URL as clean markdown for the agent to read.

Knowledge Base Search

Auto-enabled when a knowledge base is linked; returns the most relevant chunks with citations.

Knowledge Graph Search

Multi-hop Graph RAG over a knowledge base's entity graph (build the graph in Knowledge → Graph first).

SQL Query

Read-only SELECT against your CSV-derived tables, managed in Data & SQL Agents — see below.

Calculator / Date & Time / Weather

Deterministic utility tools that need no key: safe math evaluation, timezone-aware clock, Open-Meteo forecasts.

Workflow triggers

Call out to automation platforms you connect — n8n, Activepieces, Node-RED, Windmill, Temporal, Airflow, Huginn, Zapier, Make, or a custom webhook.

MCP tools

Tools exposed by Model Context Protocol servers connected at /mcp; pick which servers the agent may call.

Text-to-SQL setup

The SQL tool queries local tables created by uploading CSVs in Data & SQL Agents — column types are inferred, so you can prototype text-to-SQL agents without touching a real database. Execution is read-only (SELECT), and the platform's SQL-agent templates demonstrate the production pattern this teaches: schema-aware prompting, validation before execution, and strict scope.

Guardrails

Opt-in per agent; enabling any guardrail adds a "Guarded" badge to the agent's card.

Input Guardrails

Input filtering with a max input length and blocked patterns (regex, one per line) for prompt-injection and jailbreak strings.

Output Guardrails

Output filtering with optional hallucination detection, citation checking, and a custom filter prompt.

Rate Limiting & Boundaries

Max turns per conversation, requests-per-minute rate limit, and a token threshold above which a human approval is required.

Topic Boundaries

Allowed and restricted topic lists that keep the agent on its job.

Memory

• Short-term memory (on by default) — the conversation window, with older turns folded into a rolling summary so long chats don't lose the thread.
• Long-term memory (opt-in) — typed items (facts, preferences, episodic notes, instructions) extracted across conversations and recalled by relevance. The memory section of the form shows every stored item and lets you inspect or delete them.
• Memory tools — agents with memory enabled can call remember/recall/forget explicitly, plus a per-conversation scratchpad shared across swarm nodes.

Export, share, publish, import

• Export — LangChain (Python or TypeScript), LangGraph ReAct (Python or TypeScript), CrewAI YAML config, or a portable JSON manifest with prompts, tools, and model configuration. The generated code is readable and idiomatic — take it into your own codebase and keep going.
• Share — a link another signed-in user can import from; imports are one-time copies.
• Publish — lists the agent on the community for others to remix.
• Import — recreate any exported AgentSwarms JSON agent in your workspace.

              ┌─────────────────────────────────────────────┐
   chat ───▶  │   System prompt  +  short / long-term memory │
              └────────────────┬────────────────────────────┘
                               │
                               ▼
                        ┌────────────┐
                        │   Model    │  (provider + parameters)
                        └────┬───────┘
                             │
            ┌────────────────┼────────────────────────┐
            ▼                ▼                        ▼
        ┌────────┐    ┌─────────────┐         ┌──────────────┐
        │ Tools  │    │ Knowledge   │         │  Guardrails  │
        │+skills │    │ bases (RAG) │         │  + approvals │
        └────┬───┘    └──────┬──────┘         └──────┬───────┘
             │               │                       │
             └──────────► response ◄─────────────────┘