Route planning for AI agents: why chat assistants need real routing tools

Ask a chat assistant to put ten errands in the best order, and you will get a confident answer. But a sensible-looking order and the best order are not the same thing. This guide explains why multi-stop routing is hard for language models working alone, how AI agents get real routing ability, and what a good routing tool for an agent should do.

Why route planning is hard

Putting stops in the best order is a classic optimization problem. The best-known version is the Traveling Salesperson Problem (TSP): find the shortest route that visits every location once and returns to the start. When there are several vehicles, it becomes the vehicle routing problem (VRP).

These problems grow fast. Google’s OR-Tools documentation notes that with 10 locations (not counting the start) there are 362,880 possible routes. With 20 locations, there are about 2.4 quintillion. Checking every option is “computationally intractable for all but small sets of locations,” so real solvers use optimization techniques to find optimal or near-optimal answers.

Why chat assistants struggle on their own

A large language model (LLM) works by reasoning in text. On its own, it has no live connection to a road network, current traffic, or a distance table between your stops. It can only estimate.

Research backs this up. The GraphArena benchmark, presented at ICLR 2025, tested more than 10 LLMs on graph problems including the TSP. It sorted answers into correct, suboptimal, “hallucinatory” (properly formatted but infeasible), or missing. The authors found that even top models “struggle with larger, more complex graph problems and exhibit hallucination issues.”

A separate 2024 study of GPT-3.5 Turbo on the TSP found a similar pattern: performance was strongest on small instances, and the share of correct trips fell as the number of stops grew.

In practice, an LLM can produce a plausible route that is too long, skips a stop, or ignores a constraint.

How agents get real routing ability

The fix is to let the model call real tools instead of guessing.

Tool calling (function calling). OpenAI describes function calling, “also known as tool calling,” as a way for models “to interface with external systems and access data outside their training data.” The developer describes a tool with a schema. The model decides when to call it and with what inputs. The application runs the tool and hands the result back to the model. Anthropic documents tool use for Claude, and xAI documents function calling for Grok, so the same pattern works across these model families.

The Model Context Protocol (MCP). MCP is an open standard for connecting AI applications to outside tools and data. Anthropic introduced and open-sourced it on November 25, 2024. The official MCP site compares it to “a USB-C port for AI applications”: build a connection once, and many AI apps can use it.

In MCP, developers expose tools and data through MCP servers, and AI applications act as MCP clients that connect to those servers. The official MCP site says AI assistants like Claude and ChatGPT, plus development tools like Visual Studio Code and Cursor, support MCP.

MCP is no longer controlled by one company. On December 9, 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation.

What a good routing tool for an agent needs

A routing tool that an agent can use well should cover a few core jobs.

Geocoding. Geocoding converts an address into latitude and longitude.

Distance and time matrices. An optimizer needs the travel time and distance between every pair of stops. Google’s Routes API calculates distance and duration “for multiple origins and destinations.” OSRM’s table service computes the duration of the fastest route between all pairs of supplied coordinates.

Optimization with constraints. Real routes have rules. Common ones include:

Clean input. An optimizer is only as good as the job list it receives.

Results the agent can explain. The agent has to tell a person what happened. A useful result includes the stop order, arrival times, total distance and time, and which stops could not fit and why. GraphHopper’s documentation, for example, describes a solution that lists ordered activities with arrival and departure times, plus an “unassigned” list for orders that could not be placed.

Routing APIs and optimizers in brief

Several established tools already handle the routing math. This is a short, neutral overview, not a ranking.

Business use cases

Deliveries. Delivery teams juggle many stops, several drivers, vehicle limits, and customer time windows. Google’s documentation names logistics and on-demand delivery companies as typical users of route optimization.

Service routes. Technicians, cleaners, and maintenance crews often visit the same sites on a schedule.

The pattern is the same each time: the person states the goal, the agent structures the stops and rules, a routing tool does the math, and the agent explains the result.

Where Mappy fits

Mappy is an agent skill for route-planning workflows. It brings jobs and driver constraints into a structured request, which is the clean input this guide says an agent needs before any routing math. A Mappy route request includes a job list with addresses, service times and arrival windows, plus driver starting locations and working hours.

The Mappy site describes the skill as running through your agent, in three steps: parse jobs, resolve constraints, and plan the route. The skill package includes agent instructions, input and output contracts, a job-list template and a planning guide.

Getting started with Mappy

If you plan routes with an AI agent, you can see the Mappy agent package at mappybot.com/download. A practical first step is to put your stops into the sample job-list CSV and check them against how to prepare a job list for route planning. Your agent then has addresses, service times and arrival windows in one place.

FAQ

Can ChatGPT or Claude plan an optimal route on its own?

Not reliably for many stops. Without a routing tool, a language model has no live road distances or traffic, and research such as GraphArena shows LLMs often give suboptimal or infeasible answers on larger TSP-style problems.

What is MCP?

The Model Context Protocol is an open standard for connecting AI applications to external tools and data. Anthropic introduced it in November 2024 and donated it to the Linux Foundation’s Agentic AI Foundation in December 2025.

What is the difference between a route and an optimized route?

A route gets you from A to B through stops in a given order. An optimized route also chooses the order (and, for fleets, which vehicle visits which stop) to minimize time, distance, or cost while respecting constraints.

What constraints matter most for business routes?

Time windows, vehicle capacity, start and end locations, and driver hours are the most common. Skills, breaks, and recurring schedules also matter for many service businesses.

What is Mappy?

Mappy is an agent skill for route-planning workflows. It brings jobs and driver constraints into a structured request: a job list with addresses, service times and arrival windows, plus driver starting locations and working hours. The agent package is at mappybot.com/download.

Sources

All guides