surfs.dev Logo
surfs.dev
ResourcesNews
Brain-Body-Hands Architecture

The complete platform for
human-interface agents

Label action sequences. Train on real interactions. Deploy with safety. Eval multi-step flows. BYOB (Bring Your Own Brain).

Watch DemoStart Free Trial
3
Architecture Layers
10M+
Action Sequences
100%
Deployment Safety

Three layers, clean separation

Brain-Body-Hands architecture separates decision-making, orchestration, and execution.

agents/You control this

Brain

Your decision-making layer

OpenAI GPT-4
Anthropic Claude
Custom models
Fine-tuned agents
surfers/We manage this

Body

Our orchestration layer

Multi-step planning
State management
Error recovery
Workflow control
browser_sessions/We handle this

Hands

The execution layer

Browser automation
Desktop control
GUI interaction
Action execution
Example: Bringing your own brain
import { Surfer } from '@surfs/sdk'

// Use your own model in agents/
const agent = {
  brain: "gpt-4" // or "claude-3-opus", custom model
}

// We handle orchestration (surfers/) and execution (browser_sessions/)
await Surfer.run(agent, task)
Platform Features

Built for agent developers

Every feature designed to help you ship reliable agents faster.

agents/ - Your Brain

Bring Your Own Brain (BYOB). Use OpenAI, Anthropic, Gemini, or fine-tuned models for decision-making. You control the intelligence layer—we provide the infrastructure.

surfers/ - Our Body

Orchestration layer handles multi-step workflows, state management, error recovery, and action planning. We manage the complexity so you focus on agent logic.

browser_sessions/ - Our Hands

Execution layer for browsers, desktops, and GUIs. Browser automation with Playwright, Selenium, or native automation. We handle the physical interaction layer.

e2es/ - Validation Layer

Action-level testing framework. Validate 50/50 actions, not 49/50. Test in sandbox before production. Version control for agent behavior.

Dataset Builder

Automatically label action sequences from recorded sessions. Build training datasets for fine-tuning browser agents—not just text or images.

Deployment Pipeline

Sandbox → Staging → Production workflow. Canary rollouts for behavior changes. CI/CD for agents. Test safely before deploying to customers.

The complete agent lifecycle

From action recording to production deployment, all in one platform

Label

01

Record and annotate agent actions automatically. Capture every action in multi-step flows. Label success/failure at the action level, not just the final outcome.

Train

02

Build training datasets from labeled action sequences (not just text or images). Clean, structured data ready for fine-tuning. We're creating CommonCrawl for actions.

Deploy

03

Test in sandbox before production. Canary rollouts for behavior changes. CI/CD pipeline for agents. Never deploy untested agents to customers.

Eval

04

Validate 50/50 correct actions, not 49/50. Action-level testing catches failures that outcome testing misses. Version control for agent behavior.

Frequently asked questions

What is the Brain-Body-Hands architecture?

A clean separation of concerns: agents/ (Brain) handles decision-making with your own models, surfers/ (Body) orchestrates workflows and manages state, browser_sessions/ (Hands) executes actions in browsers, desktops, or GUIs. This architecture lets you BYOB (Bring Your Own Brain)—use OpenAI, Anthropic, or custom models for decisions.

How does BYOB (Bring Your Own Brain) work?

You control the agents/ layer—use any LLM for decision-making (GPT-4, Claude, Gemini, or fine-tuned models). We provide the surfers/ orchestration layer and browser_sessions/ execution layer. This separation means you're not locked into our model choices.

What's the difference between agents/, surfers/, and browser_sessions/?

agents/ makes decisions (your brain, your model). surfers/ plans multi-step workflows and handles orchestration (our body, manages state and flow). browser_sessions/ takes physical actions in browsers/GUIs (our hands, execution layer). Clean interfaces between each layer.

How does the Label → Train → Deploy → Eval lifecycle work?

Label: Record and annotate agent actions automatically. Train: Build datasets from labeled action sequences (not just text/images). Deploy: Test in sandbox before production with canary rollouts. Eval: Validate every action in multi-step flows (49/50 vs 50/50 correct actions). All integrated in one platform.

Build production-ready agents

Bring Your Own Brain. We handle orchestration and execution. Join 800+ developers building with Surfs.

Start BuildingWatch Demo
surfs.dev Logo
surfs.dev

The easiest way to build reliable AI agents that actually understand the web

Resources

  • Blog & Resources
  • Agentic Browser News
  • Documentation

Company

  • Privacy Policy
  • Terms of Service

© 2026 surfs.dev. All rights reserved.

Cookie Policy