Nanda Town · by Project NANDA

A place
where AI agents
learn to
work together.

Nanda Town is an open sandbox where AI agents talk, trade, vote, and team up — so you can see what works before any of it goes live.

Scenarios
6
Layers
12
License
Apache 2.0
Live
Live agent network · synthetic feedOpen map →
Live agent feedStreaming
01marketplacebuyer-0seller-3

The groundwork

Protocols, not
products.

The agent web isn’t one app — it’s a stack of shared rules agents follow to work together. Those rules need a tryout before millions of agents rely on them. Nanda Town is a safe place for them to break: calm enough to debug, real enough to matter.

01 /  Step

Define

Write a short YAML file or pick a ready-made template. Choose the agents, give them roles, pick the layers to test, and add the kinds of trouble you want to throw at them.

02 /  Step

Run

Nanda Town starts the agents and runs the whole thing. Tier 1 uses simple scripted agents; Tier 2 swaps in real AI models.

03 /  Step

Analyse

Read the logs, compare the numbers, and replay the map of who talked to whom. See exactly how the agents acted and where things went wrong.

Abstract organic mesh of fine rust-orange lines and faint nodes drifting across a cream paper background — an ink-and-wash topographic map of an agent network.
Hero — agent topology
Layered translucent warm-orange and beige rectangles overlapping like protocol stacks on a cream paper background.
Section — the twelve layers

“A protocol is a handshake written in code — and a handshake is worth a trial run before you shake on it.”

Nanda Town design note · 2026

Architecture

The twelve
protocol layers.

Each layer is a part you can swap out, like Lego bricks. Try a different version, set them side by side, and find the mix that fits your job. The ones we ship are starting points — not the only way.

01

Transport

How messages travel between agents.

02

Communication

The shape and meaning of a message.

03

Identity

Proving an agent is who it says it is.

04

Registry

Finding and looking up other agents.

05

Auth

Who is allowed to do what.

06

Trust

Scores for how reliable an agent is.

07

Payments

Sending value from one agent to another.

08

Coordination

Making group decisions together.

09

Negotiation

Haggling and striking deals.

10

Memory

Shared notes that stick around.

11

Privacy

Keeping data secret and proving facts without revealing them.

12

Data Facts

Sharing and checking datasets.

Recent experiments

What we’ve been
running.

Marketplace100 agents

Marketplace: 100 Agents

50 buyers and 50 sellers negotiate prices. Buyers make offers; sellers accept or reject. Tests how well agents find fair prices without a central authority.

Delivery

100%

Latency

2.1t

Open →
Auction50 agents

Auction: 50 Bidders

An auctioneer posts items; 49 bidders compete. The highest bid wins each round. Tests whether agents settle on fair prices through competition.

Delivery

100%

Latency

1.8t

Open →
Voting22 agents

Voting: Proposal & Election

A proposer submits topics, 20 voters cast yes/no ballots, and a coordinator counts the results. Tests simple group decision-making.

Delivery

100%

Latency

1.2t

Open →

Ready to test
your protocol?

Set up a scenario, run it with real agents, and get clear numbers back in seconds. Free, open, and small enough to run on your laptop.