Agent Framework

A standard agent asks the LLM what to do at every step, which burns tokens, adds latency, and makes behavior hard to predict. Solrouter's agent framework removes that uncertainty: each request runs through SERV (Structured Execution via Reasoning Virtualization), which replaces freeform LLM decision-making with a deterministic execution graph.

A skill-graph layer rides alongside SERV, injecting structured domain knowledge — DeFi, on-chain data, market analysis, and more — straight into the synthesis context, but only when your query actually needs it. You get an agent that reasons more reliably, spends far fewer tokens, and responds faster than a typical agent loop.

Core components

The framework rests on three pieces. Start here to learn how each one works.

Built-in tools

These are the tools every agent can reach without any setup. SERV picks which ones to call — and in what order — from the pre-defined execution graph for your query type, rather than asking the LLM at runtime.

Quick example

Here's the smallest request that exercises the whole pipeline. Send a prompt to the agent endpoint with your API key, and set useTools: true to turn on SERV-guided tool execution.

curl -X POST "https://api.solrouter.com/agent" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Compare Marginfi vs Kamino lending on Solana",
    "model": "gpt-oss:20b",
    "useTools": true
  }'

The response gives you the synthesized reply plus a record of how the agent got there: every tool call it made, how many reasoning iterations it ran, and a skill-graph summary listing which knowledge nodes it traversed.

{
  "success": true,
  "reply": "## Marginfi vs Kamino Lending Comparison\n\n...",
  "toolCalls": [
    { "tool": "web_search", "args": { "query": "Marginfi vs Kamino lending Solana" } },
    { "tool": "token_price", "args": { "token": "MNDE" } }
  ],
  "iterations": 4,
  "skillGraph": {
    "nodesTraversed": ["defi-analysis", "liquidity-risk", "comparative-analysis"],
    "relevanceScore": 0.72
  }
}