Start with the workflow
What is the intended outcome? What can the agent change? What evidence is observable?
About Loom
The model is only
part of the system.
Models are becoming increasingly capable, but the systems around them still determine what they know, what they can do, what they can observe, and when humans intervene.
Loom exists to make those systems easier to understand, evaluate, and improve.
We use harness to mean the tools, permissions, approvals, observations, and evaluation logic around an agent. A harness can prevent failure. It can also hide missing evidence, duplicate safeguards that already exist, or make the workflow impossible to complete.
Our working thesis is that a useful change must preserve both sides of the problem: protected constraints and the outcome the agent was meant to achieve.
That means examining one workflow at a time, making bounded hypotheses, and reporting uncertainty when the evidence is insufficient.
How we work
Separate evidence from inference
A benchmark result is not customer validation. A control hypothesis is not a production recommendation.
Test the trade-off
A safer-looking control is incomplete evidence if it prevents the workflow from succeeding.
Loom is early-stage research.
We are not claiming customers, production proof, or universal answers. We are looking for teams willing to discuss a specific agent workflow and how they decide it is ready.
Talk to us about your agent workflow