TL;DR: An embodied harness connects a model’s decisions to a robot’s actions. It selects sensor data, keeps track of the task, calls skills, and checks what actually happened. Meanwhile, local controllers keep the robot moving or bring it to a stop without waiting for the model.
For the past few months, we’ve been exploring agentic harnesses – it has been an especially popular topic on our YouTube channel, and we can clearly say people want even more of it. A harness is the software scaffolding wrapped around an AI model that gives it memory, tools, and guardrails to get things done. Though some argued that frontier models starting to "eat" simpler harness logic, the "model + harness" combo has proven itself as a system that really works. There are at least three reasons why:
Heavy harness infrastructure is still essential for long-horizon tasks that agents have to work on continuously for days.
Harness can be built out of modular, plug-and-play components, and there is huge potential in the idea of a harness compiling custom harnesses on demand.
And the real magic is that the exact same model can act like a completely different assistant depending on the harness it is running in.
So far, we've mostly talked about harnesses for digital agents. Today, we have numerous autonomous coding agents that can genuinely tackle long-running tasks and build software on their own. That was confirmed both at the AIE Code Summit and in our recent interview with Simon Willison. But through all of this, the topic of harnesses in robotics has barely been touched.
What if we adapt this setup to the physical world, in other words, make harness embodied and pair it with foundation models that understand 3D space, video, and physical actions?
Recently, Grid Robotics took on this challenge, laying out a blueprint for how to build agentic robotics capable of handling the same level of complexity as today's coding agents (or even higher), but in the 3D world.
Their “Envisioning Agentic Robotics: A New Era of Robotic Systems“ research caught our eye because it takes a broad look at both robotics and harness design. So let's explore it together and see what this roadmap offers on the path toward embodied agents we can finally trust to operate on their own.
Let’s go!
In today’s episode:
Why robotics needs a broader approach than VLAs and WAMs
What an embodied harness does and how it differs from a digital harness
The full workflow: What happens before and after each model call
The world doesn’t pause while the model thinks, and what happens when the user changes their mind
How to test the robot, not just the model
What makes it an embodied agent
Why do we need a broader paradigm for robotics beyond VLAs and WAMs?
Firstly, let’s turn back to the definitions of VLA and WAM that we discussed in our previous episodes.
VLAs, Vision-Language-Action models, turn what a robot sees into an action. They take camera observations, a language instruction, and sometimes the robot’s state, then predict actions.
WAMs, World Action Models, predict how the physical world may change and connect these predictions with action generation. Depending on its design, they might predict future observations alongside actions or use predicted futures to help generate them.
These model types both describe what a model does, but generalization appears to be a weak spot. In real world, either the scene rapidly changes or an instruction can be phrased differently, so performance often drops if models can't keep up with changing conditions.
But yet, VLAs and WAMs as foundation models are a promising starting point for robots that can handle a broader range of tasks. And this range is a matter of responsibility. Grid Robotics pursue creating robots which have the longer-term goal as sustained responsibility – only this way they can reliably manage workflows and recurring roles. That is a criterion of agentic robotics.
That’s why Grid Robotics’ researchers proposed a roadmap for Agentic Robotics. No, they don’t want to replace VLAs or WAMs. It is more of a gradual expansion in scope: How do we coordinate multiple actions so a robot can complete a task and verify the result?
Their idea is that a foundation model becomes the decision-making core of a robot agent when it works together with an Embodied Harness: the surrounding software that gives the model context, connects reasoning to observations, skills and tools, execution, and verification, and manages what happens when those tools act in the physical world. These two parts form a working loop of an embodied agent.
From Policy-Centric Robotics toward Agentic Robotics
Today, we mostly see policy-centric systems. In this setup, an action policy takes a task instruction and sensor readings, then generates action targets for the robot to execute – often incorporating reasoning, subtask prediction, or real-time adjustments.
In an agentic system, an agent is the decision center that decides how to pursue the goal: it can call skills and tools, check what happened, and adjust its next decision. The Grid Robotics’ researchers describe this system with three functional roles:
Layer | Role | Examples & Functions |
|---|---|---|
System 2 (S2) | High-level decision center | Interprets the user's goal, considers constraints, decides how to proceed, selects skills and tools, and checks outcomes. |
System 1 (S1) | Callable skills and tools | VLA or WAM policies, object detection, motion planning, and navigation skills. |
System 0 (S0) | Real-time body & local control | Joint movements, balance, trajectory tracking, and continuous real-time operations. |
Here, the most interested for us is S2 which combines a foundation model with an embodied harness. The idea reminds how these elements work in digital AI agent. The foundation model provides knowledge and reasoning, and the embodied harness supplies everything around: determines what information reaches the model, what it can act on, how the system follows through on its decisions, and maintains plans and memory.
We will return to how the entire agentic robot loop works later. First, we need to make clear the embodied harness’ functions →
What does the Embodied Harness do?
There are five core components that form all embodied harness jobs:
(BTW, even if you don’t work in robotics, it’s super interesting to see how an agent harness works in the physical world.)
What happens when a rescue robot gets new instructions halfway through a search? In the full guide, we break down the five parts of an embodied harness and follow a robot through a collapsed mine. See how it checks what happened after each move and changes course when the team redirects it. We’ll also look at how to test the system while the world keeps moving.
With Premium, you also get the full AI 101 archive and new deep dives in your inbox.
Previously in AI 101: The Hidden Decisions Behind an AI Answer
FAQ
What is an embodied harness?
An embodied harness is the software around a robot’s foundation model. It selects relevant sensor data, keeps track of the task, calls skills such as navigation, and checks what actually happened before the model decides what to do next.
How is an embodied harness different from a digital agent harness?
Both manage context, tools, and execution. An embodied harness also has to track where and when observations were made, work with real-time controllers, and verify physical effects from fresh sensor or controller data.
Does an embodied harness replace a VLA or a world action model?
No. A vision-language-action model or world action model can provide a skill within the larger system. The harness helps decide when to call that skill and checks whether it achieved the intended result.
What are Systems 2, 1, and 0 in agentic robotics?
In Grid Robotics’ framework, S2 is the foundation model and harness that coordinate the task. S1 supplies callable skills and tools, while S0 handles real-time control of the robot’s body.
How do you test an embodied agent?
Measure task completion separately from safety, recovery, and human intervention. Report whether the simulation kept running during model calls. To isolate the harness’s contribution, keep the model, robot, and tasks fixed and change one part of the harness.





