How to Create Spatial AI Assistants with Mistral and Harness2

Introduction: Answering the main question immediately

Spatial computing and AI fundamentals

Spatial computing transforms how devices perceive position, orientation, and geometry. When an AI assistant operates in this context, it requires the following:

• Multimodal perception (cameras, LiDAR, inertial sensors)
• Contextual awareness of the 3D world
• Response generation that accounts for spatiality

These pillars converge in modern spatial AI assistants, which can guide users through architectural layouts, overlay useful information onto augmented reality, or control robots in three-dimensional environments.

Why spatial computing is crucial today

By 2026, demand for location-based interactions has skyrocketed. New AR/VR devices integrate large language models directly on the device, eliminating the need for a constant connection. Frameworks like Harness2 enable developers to rewrite agent orchestration in real time, adapting the AI assistant's behavior to the surrounding geometry.

Choosing the right models: Mistral Timestral and beyond

When it comes to temporal and spatial reasoning, Mistral has released Timestral, a family of forecasting models ranging from 5 million to 209 million parameters. The 209 million parameter model currently dominates benchmarks such as GIFT-Eval, TIME, and fev-bench.

Timestral excels in:

• 3D time series forecasting
• Generating trajectories for robots and avatars
• Efficient inference on edge devices

Integrating Timestral into a spatial assistant

One straightforward way to leverage Timestral is through a Python wrapper:

import mistral_timestral
model = mistral_timestral.load_model("timestral-209m")

def predict_trajectory(points, horizon=10):
    return model.generate(points, steps=horizon)

Building a spatial assistant with Harness2

Google's Harness2 is an Apache-2.0 framework that rewrites agent orchestration at test time. Its dynamic adaptability makes it ideal for spatial AI assistants, where the environment's geometry changes continuously.

Key steps to integrate Harness2:

1. Define an agent skeleton with extension points.
2. Load the spatial model (Timestral) as a reasoning component.
3. Enable retrieval via TinyFish to enrich spatial data.
4. Implement a real-time feedback loop that updates orchestration based on new coordinates.

Harness2 setup example

from harness2 import AgentHarness
from mistral_timestral import TemporalForecaster
from tinyfish import WebRetriever

forecaster = TemporalForecaster("timestral-209m")
retriever = WebRetriever()
harness = AgentHarness()
harness.register_module("spatial_reasoning", forecaster)
harness.register_module("knowledge", retriever)

Prompt engineering for 3D contexts

Prompts guide the AI assistant to produce spatially aware outputs. An effective formula uses three sections:

• Context:Location, orientation, and relevant objects.
• Task:Required action (e.g., "overlay a floor plan")
• Style:Desired format (short, detailed, code, voice)

Example prompt:

Prompt:
Context: The user is in a 5m x 8m open space with a south-facing wall that receives direct sunlight.
Task: Overlay a floor plan showing the exact position of furniture and the direction of sunlight.
Style: Output in JSON format with coordinates and light notes.

Result:
{
  "floor_plan": [
    {"id":"desk","x":1.2,"y":0.8,"width":1.5,"height":0.7},
    {"id":"window","x":4.0,"y":0.0,"width":3.0,"height":2.0}
  ],
  "light_direction": "south"
}

Practical example: From user to interactive environment

Imagine you want to create an assistant that helps an architect visualize new furniture in an existing environment.

Workflow:

1. The user uploads a 3D scan of the room via the AR device.
2. The Harness2 agent extracts spatial landmarks.
3. Timestral predicts where new furniture would be positioned over time (e.g., movement during the day).
4. TinyFish retrieves up-to-date product specifications from the web.
5. The system generates a prompt asking the AI assistant to overlay the predicted furniture onto the live view.
6. The AR interface displays the results, allowing the user to rotate and resize objects with gestures.

Current trends and practical tips

• Edge device efficiency:Use Timestral-209m on spatial processing units to reduce latency.
• Code-free orchestration:Harness2 supports YAML-based configurations, accelerating time-to-market.
• Cost-free retrieval:Integrate TinyFish to enrich spatial data without research costs.
• Prompt-based iteration:Test different prompt styles (verbal vs. visual) to find the one that best suits your AR/VR hardware.

Conclusion: Create future-ready spatial AI assistants

Next steps:Choose a recent AR/VR device, install the Harness2 SDK, load a Timestral model, and experiment with a basic spatial prompt. Share your results in a comment or ask for advice on a specific use case!

Introduction: answer the main question immediately

Fundamentals of spatial computing and AI

Spatial computing transforms how devices perceive position, orientation, and geometry. When an AI assistant operates in this context, it needs:

• Multimodal perception (cameras, LiDAR, inertial sensors)
• Contextual awareness of the 3D world
• Response generation that takes spatiality into account

These pillars come together in modern spatial AI assistants, which can guide users through architectural layouts, overlay useful information onto augmented reality, or control robots in three-dimensional environments.

Why spatial computing is crucial today

By 2026, demand for location-based interactions has exploded. New AR/VR devices integrate large language models directly on the device, eliminating the need for a constant connection. Frameworks like Harness2 allow developers to rewrite agent orchestration in real time, adapting the AI assistant's behavior to the surrounding geometry.

Choosing the right models: Mistral Timestral and beyond

When it comes to temporal and spatial reasoning, Mistral has released Timestral, a family of forecasting models ranging from 5 million to 209 million parameters. The 209 million parameter model currently dominates benchmarks such as GIFT-Eval, TIME, and fev-bench.

Timestral excels in:

• 3D time series forecasting
• Generating trajectories for robots and avatars
• Efficient inference on edge devices

Integrating Timestral into a spatial assistant

One simple way to leverage Timestral is through a Python wrapper:

import mistral_timestral
model = mistral_timestral.load_model("timestral-209m")

def predict_trajectory(points, horizon=10):
    return model.generate(points, steps=horizon)

Building a spatial assistant with Harness2

Google's Harness2 is an Apache-2.0 framework that rewrites agent orchestration at test time. Its dynamic adaptability makes it ideal for spatial AI assistants, where the environment's geometry changes continuously.

Key steps to integrate Harness2:

1. Define an agent skeleton with extension points.
2. Load the spatial model (Timestral) as a reasoning component.
3. Enable retrieval via TinyFish to enrich spatial data.
4. Implement a real-time feedback loop that updates orchestration based on new coordinates.

Harness2 setup example

from harness2 import AgentHarness
from mistral_timestral import TemporalForecaster
from tinyfish import WebRetriever

forecaster = TemporalForecaster("timestral-209m")
retriever = WebRetriever()
harness = AgentHarness()
harness.register_module("spatial_reasoning", forecaster)
harness.register_module("knowledge", retriever)

Prompt engineering for 3D contexts

Prompts guide the AI assistant to produce spatially aware outputs. An effective formula uses three sections:

• Context:Position, orientation, and relevant objects.
• Task:Required action (e.g. "overlay a floor plan")
• Style:Desired format (short, detailed, code, voice)

Example prompt:

Prompt:
Context: The user is in a 5m x 8m open space with a south-facing wall that receives direct sunlight.
Task: Overlay a floor plan showing the exact position of furniture and the direction of sunlight.
Style: Output in JSON format with coordinates and light notes.

Result:
{
  "floor_plan": [
    {"id":"desk","x":1.2,"y":0.8,"width":1.5,"height":0.7},
    {"id":"window","x":4.0,"y":0.0,"width":3.0,"height":2.0}
  ],
  "light_direction": "south"
}

Practical example: From user to interactive environment

Imagine you want to create an assistant that helps an architect visualize new furniture in an existing environment.

Workflow:

1. The user uploads a 3D scan of the room via the AR device.
2. The Harness2 agent extracts spatial landmarks.
3. Timestral predicts where new furniture would be positioned over time (e.g. movement during the day).
4. T