SERV Reasoning

A standard agent loop asks the LLM what to do at every step, which is slow, expensive, and unpredictable. SERV (Structured Execution via Reasoning Virtualization) fixes that by separating the two jobs an agent actually does: deciding what to gather, and writing the answer.

Instead of asking the model what to do next, SERV walks a Guided Reasoning Diagram (GRD) — a pre-defined execution graph that controls tool selection, sequencing, and data collection on its own, with no LLM in the loop. The model is called exactly once, at the very end, to turn the collected data into a natural language response.

Splitting structure from synthesis is the whole trick: it makes SERV far cheaper and faster than a standard agent loop while keeping output quality high.

Performance results

Here is what that split buys you, benchmarked against a standard agent loop on identical complex research queries:

SERV cuts cost and latency and improves quality, with no loss in reliability. The quality gain is not magic: because the GRD gathers data the same structured way every time, the model synthesizes from a complete, consistent picture instead of improvising its own research path.

How SERV works

This section walks through the four stages a query passes through, and why each one keeps the LLM out of the decisions it isn't good at.

A standard agent loop asks the model the same question at every step: "Given what you know so far, what tool should you call next?" That burns tokens on structural choices the model has no special advantage in making, and it introduces non-determinism that can derail a complex query halfway through.

SERV takes a different approach:

The key insight: don't ask a 20B model to make structural decisions. Do those deterministically, and reserve the LLM for the one task it does best — turning raw information into useful language.

Using SERV in the SDK

The fastest way to try SERV is through the SDK. Pass reasoning: 'braid' to the chat() method, and your request routes through the agent endpoint with SERV-guided reasoning enabled.

import { SolRouter } from '@solrouter/sdk';

const client = new SolRouter({
  apiKey: 'sk_solrouter_...'
});

const response = await client.chat('Compare Marginfi vs Kamino lending on Solana', {
  reasoning: 'braid',  // enables SERV-guided reasoning
});

console.log(response.message);

Using SERV via the API

If you're not using the SDK, you can call the agent endpoint directly over HTTP. Set useTools: true to activate the SERV execution graph.

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 includes an iterations field showing how many GRD steps SERV executed, and a toolCalls array logging every tool the agent invoked and with what arguments.