What's in a Harness?
A bank asked us what was in our harness. Here is our answer.
A bank asked us recently what was in our harness. Fair question. It is also one of those words AI people have started using as if everybody agreed on a definition.
For us, the harness is the software around the model that lets it carry out an investigation: the instructions it follows, the tools it can use, the evidence it keeps, and the points where it needs a person's help.
Take a business with an unclear ownership structure. The model needs to find the relevant records, read the documents, work through the ownership chain, and identify what is still missing. It also needs to know the bank's policy. A plausible answer from memory is not much use here.
We think about the system in a few connected parts:
| Part | What It Contributes |
|---|---|
| Data and research | Business records across all 50 states, federal sources, sanctions, adverse media, and relevant public web and social information |
| Document extraction | Usable information from filings, ownership documents, and the PDFs that arrive with a case |
| Agent runtime | OpenClaw-based infrastructure for carrying out research and coordinating tasks |
| Model routing | Choosing models for different jobs, balancing quality, speed, and cost |
| Policies and review | Customer instructions, evidence requirements, escalation points, and analyst oversight |
| Memory | Retrieving relevant knowledge from earlier work and preserving useful corrections |
These parts have to work together. Finding an article is useful only if we can establish that it concerns the right person. Extracting an ownership percentage helps only if we understand where it sits in the chain. A policy change needs to reach the agent doing the work.
A lot of our engineering lives in those details.
The part we are especially excited about is memory. We already use it internally to help run Sweat. The visualization looks like a ball of knowledge: every dot represents a note, with related subjects forming clusters.
Underneath that picture, vector representations help us find related information by meaning. Similarity helps rank what might be relevant; it does not establish whether a claim is true. The source and the circumstances still matter.
We are building toward using the same approach in customer investigations. An analyst's correction, a previously resolved ambiguity, or a customer-specific interpretation of policy should be available when a similar question comes up again. That requires careful control over whose knowledge is retrieved and whether it is still current.
So what is proprietary? Our investigation workflows, the way we connect evidence to policy, and the integration that brings research, memory, and human review into the same case. We also use third-party tools and open-source infrastructure, and have open-sourced part of our own memory work.
The practical test is fairly ordinary: when an analyst opens a case, has the system done useful work, can they check it, and does it remember what they corrected last time?