The Harness

The core idea

You Already Use a Harness

Open any modern AI assistant and ask it to help you plan a dinner party for this weekend. Watch what happens.

The model does not just think and reply. It searches the web for trending recipes. It checks the weather forecast to decide whether you should plan for the patio or the dining room. If you told it last month that your sister is vegetarian, it remembers and adjusts the menu. It finds a playlist on a music service. It drafts a shopping list and offers to email it to you.

That feels like one seamless interaction, but underneath it several distinct things are happening. The model is the language intelligence: it understands your request and generates a response. But everything else (the web search, the weather check, the email draft) are tools the model was permitted to call. The fact that it remembered your sister's dietary restriction from a previous conversation is a memory. The way all of this information was gathered and assembled into a single coherent answer is context. And the code that decides when to search, what to remember, which tools to invoke, and how to present the result is the harness.

In personal AI, this works remarkably well. The data is simple, the stakes are low, and the only person who needs to trust the output is you. But what happens when the AI needs to work not with public knowledge and personal preferences, but with a firm's live, sensitive, fast-changing enterprise data?

Why it breaks

Why This Breaks in Enterprise Finance

In a companion piece, The Stage Is Set: Why Enterprise AI in Finance Is No Longer a Future State, I examined why the gap between personal AI and enterprise AI is not a technological problem but an architectural one. The harness is where that architecture lives, and it is where personal AI and enterprise AI diverge most sharply. Three fundamental challenges explain why.

These three challenges (retrieval of live data, preparation of meaningful context, and enforcement of governance) are not features you add to a general-purpose AI platform. They are the core responsibilities of the harness itself.

The industry consensus

Agent = Model + Harness

The idea that the code wrapped around an AI model matters as much as the model itself has rapidly become a point of convergence across the AI engineering community. What was recently an intuition among practitioners is now a documented, measured discipline that even has a name: harness engineering.

Harrison Chase, co-founder of LangChain, has argued that better models alone will not get an AI agent to production, and has traced the evolution from simple chains to complex orchestration flows to full agent harnesses as the field's true frontier. Birgitta Böckeler, a Distinguished Engineer at Thoughtworks, formalized the concept in February 2026 with her “guides and sensors” framework: guides that direct agent behavior and sensors that feed back what the agent is doing. In the same month, OpenAI described how a small team of three engineers drove its Codex agent to produce on the order of a million lines of code by changing nothing but the harness, distilling the philosophy into three words: “Humans steer. Agents execute.” And in March 2026, Stanford and MIT researchers published the Meta-Harness paper, demonstrating that automatically optimizing the harness around a fixed model can outperform hand-engineered alternatives and, critically, that a good harness transfers across different models.

A harness is not just a configuration or a set of prompts. It is a stateful program that determines what information to store, retrieve, and present to the model at each step. The model reads text and produces text. The harness decides everything else. The diagram below shows the full anatomy: the model sits at the center, and the harness is the governed system that surrounds it and connects it to the firm.

Model quality has become table stakes. The harness is the differentiator. For general-purpose AI this insight matters. For regulated finance, it is everything.

A concrete scenario

The Advisor's Reality

To make the building blocks concrete, consider a scenario that plays out at wealth management firms every day. A financial advisor is preparing for a quarterly client review. Today, she logs into four or five separate platforms: the portfolio accounting system for current positions, the CRM for client notes and life events, a market research portal for macro views, a financial planning tool for projections, and a document management system for prior meeting notes and investment policy statements. She spends an hour assembling a picture that should take minutes.

An AI agent with the right harness changes this entirely. But the harness must be built specifically for this world. Its four building blocks, memory, context, capabilities, and guardrails, each take on a fundamentally different character in enterprise finance.

In personal AI, memory means the model's recall of your past conversations: that you prefer window seats, that you asked about hiking trails last week. In the AI engineering community, this has a well-established meaning: short-term memory (the current conversation) and long-term memory (preferences and patterns that persist across sessions). Both matter.

But in enterprise finance, a third and dominant category exists: institutional memory. When our advisor sits down to prepare for her client review, the memory she needs is not a summary of her last AI conversation. She needs current portfolio positions, recent transactions, risk exposures, life events that affect the financial plan, regulatory flags, and market movements relevant to the client's holdings. None of these are in the conversation, and if they are, they are already stale.

That is the real memory. It is live. It is institutional. And it does not belong to the model; it belongs to the firm. Portfolio positions updated in real time via SQL queries against live systems. Client records retrieved from CRM platforms. Deal documents surfaced from vector stores. Connected, not copied, not cached, not stored inside a model provider's infrastructure.

There is another dimension that general-purpose platforms miss. Consumer AI memory is personal: one user, one history. Financial services memory is team- and role-level. Our advisor shares a client book with her team. A deal team shares pipeline intelligence. A risk team shares exposure data across the portfolio. When a team member leaves, the memory does not go with her.

In personal AI, context is simple: your question, plus whatever the model found on the web, plus what it remembers about you. The harness assembles the prompt, and the model responds.

For our advisor, the context challenge is entirely different. The harness needs to assemble her client's current positions from the portfolio accounting system, recent life events from the CRM, the investment policy statement from the document repository, relevant market commentary, and prior meeting notes, all scoped to what this advisor, on this team, for this specific client, is authorized to access. Show the model too little, and it hallucinates. Show it too much, and the signal drowns in noise. Show it the wrong thing, and you get a perfectly articulate, completely misleading output.

The context window is not “what does this individual user need?” but “what does this team's mandate require for this question?”

Context in finance is team-shaped, not user-shaped. An advisory team, a credit team, and a compliance team need different slices of the same enterprise data. Teams also develop their own language, shorthand, and calculation conventions that carry deep meaning but that no LLM has been trained on. Encoding this institutional context into the harness is what makes the difference between AI that sounds knowledgeable and AI that actually is.

In personal AI, tools are open-ended: the model searches the web, executes code, creates files, and sends emails with minimal constraints. In enterprise finance, every capability is governed, scoped to what this team is allowed to do, logged for auditability, and integrated into workflows that include human judgment.

The capability layer of an enterprise finance harness works across four layers, grouped into two categories. The first two form the context infrastructure that prepares what the model sees. The other two are tools in the technical sense: actions the model invokes during a conversation. They build on each other, from the foundational data work at the bottom to the point-of-use tools at the top.

Data Augmentation · Context infrastructure

This is the most valuable and hardest-to-replicate layer. It creates information that does not exist in the raw data. When our advisor asks why her client's portfolio underperformed its benchmark, the answer requires performance attribution, decomposing returns into allocation, selection, and interaction effects. No LLM can compute this reliably in real time. The harness provides this derived intelligence as a pre-computed, governed context. This is the layer that answers the question every firm eventually asks: “Why can't we just plug an LLM into our data warehouse?” Because the warehouse does not have what the advisor actually needs. Someone has to compute it. That is part of the harness.

AI Data Readiness · Context infrastructure

Data engineering moves data. AI data readiness makes data comprehensible to an AI operating within a specific team's domain. It includes data dictionaries that tell the AI what each field means in this team's context, profiling so the AI understands data quality before it answers, and team-level linguistic context: the terms, acronyms, and calculation logic that an advisor uses daily, but no LLM has been trained on. These are institutional definitions, the linguistic DNA of how a team thinks about its work.

Workflow Orchestration · Tool

The AI helps our advisor prepare a client review. Then what? In enterprise finance, the answer is rarely “just show it on screen.” It is “route the draft for compliance review,” “flag the allocation change for the investment committee,” “generate the meeting summary and send it to the client.” The harness needs tools that trigger structured workflows with human checkpoints at the points where judgment matters. Each step logged. Each decision traceable.

Productivity Integration · Tool

Our advisor does not live inside an AI interface. She lives in email, Teams, calendars, and documents. The harness needs tools that reach into these environments: pull context from a recent email thread with the client, surface the last meeting's action items, draft a follow-up, and send it, all within the same governance framework.

In personal AI, guardrails are basic safety filters. In regulated finance, they are the foundation for everything else, and they are not one thing. They are three distinct disciplines.

Security controls who can see and do what. When our advisor's AI agent queries the portfolio accounting system, the identity context must flow through every layer. The system needs to know that this specific advisor, on this specific team, initiated this specific request. Most enterprise AI implementations quietly break here: they connect the AI to source systems using a service account with broad access, bypassing the entitlement controls the firm spent years building. The security boundary must be at the point of retrieval, not the point of display.

Governance controls quality. Where did this data originate? When was it last validated? Is it complete? In financial services, a wrong number is not an inconvenience; it is a regulatory issue. The harness must embed quality controls at the retrieval layer, adding assurance at the point where information is actually consumed.

Compliance controls auditability. Every interaction must be reconstructable for audit, regulatory inquiry, or internal review. Every data retrieval is logged with full provenance. Every output is traceable end to end. The complexity of the compliance path should match the stakes of the decision: a simple portfolio lookup follows one path; a client recommendation follows a different one, with draft, cross-reference, human review, and then delivery.

These three disciplines are not a tax on the harness. They are its most valuable output, and the most secure, well-governed, compliant architecture is also the fastest. When controls are embedded in the harness rather than applied as separate review cycles, the review is the system.

The strategic stakes

Who Controls Your Harness?

The central argument of harness engineering deserves repeating in the context of financial services: because memory lives inside the harness, whoever controls the harness controls the memory. In regulated finance, the memory is the firm's most sensitive asset: client data, portfolio intelligence, deal pipelines, and risk exposures.

If your harness is embedded in a vendor's proprietary platform, the vendor controls it. Your memory, your context, your guardrails, your audit trail, all on their infrastructure, under their terms.

The model, by contrast, can be swapped. Models improve every quarter. The Stanford Meta-Harness research confirms this: harness optimization transfers across models, and a well-designed harness can make a smaller, cheaper model outperform a larger one running on a generic scaffold. The harness cannot be swapped easily. It encodes how your firm thinks, how your teams operate, what your compliance framework requires, and how your data flows.

No two firms should have the same harness, because no two firms have the same context.

The Clarista view

Built, Not Imagined

At Clarista, this is not a position paper; it is the architecture we have built and continue to evolve. Our architecture treats the firm's live, governed data as the agent's institutional memory, retrieved in real time, never copied into a model provider's infrastructure. Context is assembled at the team level. Our capability layer computes derived financial intelligence, makes data AI-ready through dictionaries and team-level linguistic context, orchestrates governed workflows, and integrates with the productivity tools where teams actually operate.

Security, governance, and compliance are not features we added; they are the foundation we built on. The authentication chain flows from the user through to every data source. Entitlements are enforced at the retrieval layer. Every output is traceable from source to response.

The model provides the intelligence. The harness provides trust, compliance, and a competitive advantage.

The firms that will define the next era of financial services are making harness decisions right now, whether they use that language or not. The ones who recognize this will own their advantage. The ones who do not will rent it, on someone else's terms.

The stage was set. The harness is how you step onto it.

Evidence

Sources & References