Node reference
An agent is a graph of nodes. There are six types.
Input
The entry point. Whatever you send when testing — or in the input field of an
API call — arrives here and flows to the connected nodes.
Agent
An LLM step. Configure the provider and model, plus a system prompt.
Agent nodes inherit the agent's compiled rules (see below) and can add a node-specific appendix in their own prompt field. This is where tool calling happens: an agent node connected to tool nodes can decide to call them.
Tool
Calls either a built-in tool or your own Python.
Built-in tools (1 credit per call):
| Tool | What it does |
|---|---|
| HTTP Request | Call an external HTTP endpoint |
| Web Scrape | Fetch and extract page content |
| JSON Parse | Parse a JSON string into structured data |
| JSON Stringify | Serialize data to a JSON string |
| Text Split | Split text into chunks |
| Text Join | Join text fragments |
| Send Email | Send a transactional email |
| Date & Time | Current date/time and formatting |
| Delay | Pause execution for a set duration |
Custom Python tools (5 credits per call) run in a Pyodide WebAssembly
sandbox with a hard 15-second timeout and globals reset between runs. Dangerous
primitives (subprocess, ctypes, raw sockets, eval/exec) are blocked
before execution, though HTTP via urllib is available.
See Tools for the full sandbox rules.
Condition
Branches the graph. Evaluate an expression and route execution down different edges depending on the result — used for quality gates, retry loops, and routing between specialist agents.
Retrieval
Queries a vector database and returns matching chunks for a downstream agent node to ground on.
Supported backends:
- pgvector (Postgres)
- Pinecone
- Qdrant
- Weaviate
Query embedding works in two modes: backend sends the raw text and lets the
vector store embed it, or byok embeds the query with your own OpenAI or
Gemini key and sends the vector. Connection configs are encrypted and decrypted
only inside the worker.
Output
Captures the final result of the run. What reaches this node is what the API returns and what appears in the test panel.
Agent rules
Beyond per-node prompts, each agent has a set of typed rules — identity, mission, prohibitions, tone, edge cases — authored under Dashboard → Agents → <agent> → Rules.
Rules compile deterministically into the system prompt every LLM node inherits. They can be toggled and versioned individually, and an on-demand conflict analysis flags rules that contradict each other — the failure mode that silently degrades agents in production.