Butterpath

Concept

DeepAgent Food Tours is a multi-agent system that plans personalized food tours from natural language. Instead of manually researching restaurants and neighborhoods, you draw search areas on a map, type something like “fun evening exploring local food,” and the agents handle the rest: searching establishments, analyzing neighborhood culture, and generating an HTML dashboard.

Built with LangChain DeepAgents, Google Maps Platform, and the Tavily research API. The architecture is a test of whether specialized agents can self-orchestrate a complex planning task without explicit flow control.

This started as a technical prototype for Butterpath, an audio food tour concept that turns city walks into curated culinary routes. The goal is the opposite of generic tour apps: authentic, self-guided experiences that feel like walking with a friend who has lived there fifteen years and knows which banh mi shop has the infamous fly story.

Architecture

Three coordinated services power the system:

• Scan Manager: Interactive map interface where users add search points and configure scan parameters. Performs lightweight Google Places searches with automatic deduplication, manages the DeepAgent API bridge, and handles user selection before expensive detailed API calls.
• DeepAgent API: Core orchestration using LangChain DeepAgents with three specialized agents:
  • Restaurant Finder Agent: Searches for establishments matching the prompts, evaluates reviews and photos, selects optimal candidates
  • Neighborhood Researcher Agent: Analyzes local food culture, trends, and context using Tavily research Dashboard Creator Agent: Generates HTML tour reports with establishment details, reviews, and personalized recommendations Dashboard Server: Serves auto-generated HTML tour dashboards with photos, reviews, and interactive route visualization. Agent Coordination: Uses LangGraph’s SubAgentMiddleware for task delegation and TodoListMiddleware for planning. The main agent breaks down “plan a romantic dinner date” into discrete tasks, delegates to specialized sub-agents, and synthesizes results. This self-orchestration pattern eliminates manual flow control while maintaining reliability.

Key Learnings

• DeepAgents’ planning middleware works for this shape of task. TodoListMiddleware lets the main agent decompose “plan a romantic tour” into discrete tasks (research neighborhood, search Italian restaurants, find dessert spots, optimize route) and delegate to sub-agents. No hand-written sequencing logic.
• Multi-point scanning matters. Single-point searches miss establishments at the edges of an area. Scanning overlapping circles across neighborhoods and deduplicating results gets coverage close to full for relevant establishments.

Status

Launched November 6, 2025. Currently supports basic scanning and AI-powered tour planning worldwide. Near-term roadmap: route optimization, dietary restrictions, real-time availability.

The longer-term direction is location-aware audio tours that auto-start when you arrive, pause when you remove headphones, and remember why you skipped that tapas place — a compounding memory system for urban culinary exploration.