#Custom Agents

Beyond the builtin agent types, you can create custom agents by composing reusable components into a graph workflow.

#Graph Model

Agents in SciLaxy are directed graphs where:

• Nodes represent processing steps (LLM calls, tool executions, decision points)
• Edges define the flow between nodes
• Conditional edges enable branching based on node outputs

The graph is compiled into a LangGraph StateGraph that manages state transitions and execution.

#Components

Components are reusable subgraphs that handle specific capabilities:

Components are registered globally and resolved by name at compilation time. You can create new components and use them in any agent graph.

#Agent Configuration

An agent's graph is defined in a JSON configuration:

{
  "agent_type": "custom",
  "graph": {
    "nodes": [
      {
        "id": "research",
        "component": "react",
        "config": {
          "tools": ["web_search", "literature_search"],
          "system_prompt": "You are a research assistant."
        }
      },
      {
        "id": "summarize",
        "component": "react",
        "config": {
          "tools": [],
          "system_prompt": "Summarize the research findings."
        }
      }
    ],
    "edges": [
      { "from": "START", "to": "research" },
      { "from": "research", "to": "summarize" },
      { "from": "summarize", "to": "END" }
    ]
  }
}

#Node Configuration

Each node specifies:

• id — Unique identifier within the graph
• component — Which reusable component to use
• config — Component-specific configuration (tools, prompts, etc.)

#Edge Types

• Direct edge — Always transitions from one node to another
• Conditional edge — Evaluates a condition to choose the next node

#Compilation Pipeline

Agent configs go through a four-stage pipeline before execution:

Agent Config (JSON)
    ↓ canonicalize
Canonical Config (normalized, deterministic)
    ↓ validate
Validated Config (structural integrity verified)
    ↓ compile
LangGraph StateGraph (executable workflow)
    ↓ .compile()
Runnable (ready to stream)

1. Canonicalize — Normalize the config format, resolve component references, apply defaults
2. Validate — Check that all nodes are reachable, edges are valid, and tool references exist
3. Compile — Build LangGraph nodes and edges from the canonical config
4. Execute — Run the compiled graph with streaming callbacks

#Example: Research Agent

For example, a custom research agent can be composed from components:

START → clarify → brief → supervisor → report → END
                    ↑         │
                    └─────────┘  (if quality insufficient)

• clarify — Breaks the research question into sub-questions
• brief — Runs web search and literature search in parallel
• supervisor — Evaluates gathered content, may loop back to brief
• report — Synthesizes everything into a structured report

Each node uses a different component with specialized prompts and tools. The supervisor node uses a conditional edge that either proceeds to report or loops back to brief for more research.