
What is an AI harness, and why does it matter for private funds?
An AI harness is everything around a model that gives it instructions, tools, context and checks so it can perform a task. In a private fund, an agent harness connects the model to fund terms and a calculation engine, with sources and human approval required before a result posts.
Your controller gets the fee with its source attached, and nothing posts until someone on the team approves it.
The model reads the request. The harness tells it what the job is, gives it the documents and systems it needs, and decides what has to happen before its work counts.

In the LP 14 case, the instruction is the fee question, the side letter is the context, the calculation engine is the tool and the controller's approval is the check.
The word comes from horses. A horse has strength, but a harness connects that strength to a cart and lets the driver steer it.
An AI model is powerful in the same way. On its own, it can't see the fund's data, use its systems or check a result against the fund's records.
Everyone talks about AI models. But the model is only half of the answer. The other half is the harness. The word comes from horses. A horse is strong. But without a harness, it cannot pull a cart, and you cannot steer it. An AI model is the same. It is powerful. But on its own, it cannot see your data, use your systems, or check its work. The harness is everything around the model. Instructions tell it the job. Tools let it act. Context gives it the facts. And checks stop it before it makes a mistake. Same model, different harness, very different results. In a fund, this matters. Ask a model alone for the fees on one LP, and it guesses. Put it in a harness, and it reads the fee terms from the side letter, runs the calculation in a system built for it, shows the source, and waits for your team to approve. Maybern builds that harness for fund operations. Your fund terms become rules. One engine calculates each number. And agents work inside those rules, with your team in control. Maybern. Get the harness right.
“Same model, different harness, very different results.” What the model can answer depends on what it can see, what it can call and what has to approve the result. Without LP 14's side letter it falls back on what fees usually are, and 1.5% is a plausible wrong answer. With clause 4 it has the rate, and with the engine it never does the arithmetic itself.
“What management fee does LP 14 pay?”
In this illustrative example, a model alone guesses “1.5%?” In a harness, it reads clause 4 of LP 14's side letter, which sets a 1.25% annual management fee on a $3.3M commitment.

The model calls a calculation engine. The engine applies the configured fee rule:
$3.3M × 1.25% = $41,250 annually.
The model does not do the math. Billed quarterly, that is $10,312.50 a quarter. The agent presents the calculation with its source and waits for the team to approve.
The controller can review the rate against clause 4 and confirm the commitment used as the fee basis. Each posting needs its own approval, so next quarter's fee comes back to the controller with its own source.
Context is the information available to an AI tool when it answers. It is one part of the harness. Context supplies the facts, while tools let the model request calculations or perform other permitted actions.
A context graph is the connected record of decisions behind those numbers: who approved what, under which rule and why. A harness reads that record before an agent acts, and each approval adds to it.
Maybern encodes LPA terms as rules, applies them to transactions and maintains the firm's calculated fund record. A side-letter rate, like LP 14's 1.25%, is set on that investor's commitment. One engine calculates each result and links it to its source.
The fund record supplies the rules and calculation engine, with an approval step before posting. Maybern's agents work inside those controls. Nothing posts until a person approves, and agent actions are logged.
Your AI tool connects to the fund record through the Maybern MCP, or Model Context Protocol. It requests calculations from the engine and receives results with their sources. Your controller reviews the same numbers, from the same fund record, that your AI tool used.
Instructions can take the form of a system prompt. Tools can be defined functions or MCP tools, with context drawn from the fund record. Permissions govern available actions. Approval gates control posting, and an action log records what the agent did.
The work of configuring and testing these components is often called harness engineering. The reference architecture shows how Maybern's results and sources reach the firm's own Snowflake account.
Next in AI for Private Funds: tool calling.
Quick reference for this topic.
For a fund question, the harness decides more. On its own, a model guesses 1.5% for LP 14. In a harness, the same model reads clause 4 and gets a $41,250 calculation from the engine. The harness also holds the result for a person's approval before anything posts.
An AI agent uses a model and tools to carry out a task. The harness is the surrounding structure that supplies its instructions, context and controls. It defines what the agent can use and when work requires approval.
Yes, if it is connected to the fee terms and a calculation engine. Without them, any model is estimating. For a fund fee, the rate should come from the side letter and the arithmetic from an engine. The result goes to a person for approval with its source attached.
The reviewer should confirm the applicable terms and calculation inputs, including the effective date. The result should link to its source and calculation method. A person must approve before posting, and the workflow should log the agent's actions.



