Building Company Intelligence: Private Evals, Open Harnesses, and Agents That Run the Work

How does a company create AI of its own?

A company can use a powerful model and still need a way to improve it for its own work. That is the opening problem in the No Priors × Latent Space conversation at Microsoft Build: what stack and tooling let enterprises create intelligence of their own while using other providers’ AI? The platform strategy centers on customers building differentiated intelligence.

The proposed foundation for Microsoft’s MAI models is clean model lineage, high-quality pretraining data, and ablations. The respondent argues that some open-weight models perform well on a few benchmarks without being equally useful in practice. The companion’s MAI technical takeaways provide a related written discussion.

A specialization loop then turns the generalist toward company-specific work: surround the model with an iterative improvement system, collect traces—records of its work—and measure progress against private evaluations. These evaluations test the objectives that matter to the company. The argument is that they become more consequential as public benchmark scores approach saturation.

A customer demo illustrates the proposal: collect traces using another model, then obtain a higher result with a five-billion-parameter reasoning model. Specialization can improve a smaller model’s task performance over time. The reported result comes without a task definition, metric, score, training procedure, or comparison conditions; it does not establish a general small-model advantage.

Agent throughput creates a human review bottleneck

Reflecting on the preceding few years, the respondent says increased compute has delivered capability gains, while the industry underestimated the difficulty of turning those capabilities into useful workflows. The ultimate evaluation is whether users accomplish valuable, measurable tasks.

Coding exposes the difficulty. A developer managing a hundred agent sessions can inherit excessive cognitive work: understanding what happened and deciding what to do next. This creates a need to rebuild the integrated development environment, or IDE. Chat alone is insufficient; a canvas and other interfaces are needed to make the work comprehensible.

The same pattern could extend to coordination and work requiring human judgment. Long-running agents could operate overnight with delegated authority, shortening workflows and completing tasks. The six-month prediction made in this Build conversation also entails new interfaces for reviewing delegated work. More work can happen while a person is away, but that person still needs to inspect what happened on their behalf.

The harness connects models, context, and tools

An enterprise harness is the surrounding system that connects models, data, and tools in a working loop. Microsoft’s products are described as using multiple models and exposing tools progressively to improve token efficiency. Preparing rich context—the information an agent needs for its task—is a major determinant of whether its plan executes efficiently.

The respondent says Microsoft uses a GitHub harness across its products and makes it available in Foundry. Customers can also bring their own or other open harnesses, together with their tools, context, and multiple models. Improvement therefore depends on the surrounding system as well as the model.

A security system, transcribed uncertainly as “MDASH,” reportedly found bugs or vulnerabilities that Mythos had not found. The respondent offers this as evidence that a harness combining models can improve real-world performance. No vulnerability counts, test scope, model versions, or access conditions accompany the comparison, so it remains an example of possible system-level gains rather than a controlled performance result.

Architecture leads directly to ownership. A platform provider can have successful first-party products, the respondent argues, while leaving room for customers to achieve comparable success. Three assets matter to that control: - Private evaluations encode the objectives against which a company improves its system. - Protected traces preserve experience gained while pursuing those objectives without leaking it. - An open harness, context, and tools let the improvement process continue across underlying models.

The practical test is whether a company can replace model A with model B and keep improving against its private evaluation. That ability would let customers retain control and compound differentiated value as the underlying platform improves. In this account, enabling others to build their own intelligence layers is what makes the offering a developer platform.

Work traces could train a company veteran

The next question is how company-specific intelligence changes intellectual property. Human judgment remains central to the response: people identify gaps and opportunities. The challenge is to combine that agency with growing access to tokens so that the two compound.

Consider the proposed Teams example: humans and agents work together, producing traces of how the enterprise creates value. Those traces could train a company veteran agent with accumulated organizational experience. Its learning resource would be the history of people and agents getting work done.

The respondent speculates that this captured expertise could eventually deserve a place on a company’s balance sheet—a proposed future, not an established accounting treatment. The underlying idea is that learning from accumulated traces could make tacit organizational knowledge more durable.

Unbundle SaaS while preserving its business knowledge

The discussion turns to enterprises generating workflows they previously bought as software as a service, or SaaS. The conventional application bundles a business-process data model, business logic, and a user interface. Agents create an opportunity to repackage those layers while retaining useful work already embedded in them.

Two examples distinguish durable assets from packaging that can change: - Stable schemas remain useful. A general ledger benefits from robust entity relationships; generating a new application does not create a reason to reinvent them. - Business logic can outlive its dashboard. Power BI’s underlying semantic models and measures capture business definitions and calculations that remain valuable when exposed for new uses.

Work IQ exposes organizational information beyond its original applications. It is described as opening access to data previously reached principally through Microsoft 365 applications, including email, Teams, Word, Excel, PowerPoint, and SharePoint.

The reported workflow starts inside a GitHub repository: ask for relevant design meetings from the previous week, then ask what code changes should follow. Work IQ can consult the meeting transcripts and return a plan for changing the codebase. Meeting context becomes input to a code-change plan, connecting organizational discussion to implementation.

Serving those agents also changes infrastructure demand. The respondent predicts that agent usage of Microsoft 365 could exceed end-user usage and says systems built to serve inboxes or mailboxes need architectural changes. The accompanying estimate of a tenfold value opportunity has no specified baseline or measurement method.

Meter agent work and include maintenance in build-or-buy

Different pricing models address different buyer preferences: - Per-user subscriptions provide budget certainty by bundling usage entitlements. - Consumption pricing meters variable work, including agent activity beyond a person’s interactive usage. - Outcome pricing shares in the result, which some customers find less attractive once a valuable outcome makes the payment feel like a royalty.

GitHub Copilot illustrates why a meter matters. Its original per-user design assumed interactive code completion and tasks. A hypothetical user launching ten thousand agents that run all day creates a different consumption pattern. The respondent expects per-user pricing to persist alongside a consumption meter.

For enterprises considering whether to rebuild SaaS tools, the respondent expects at least a full budget cycle to reveal the equilibrium. The decision rule includes creation and upkeep: buy when building and maintaining internally costs more. Generating software still has marginal costs.

Maintenance makes that tradeoff concrete. AI may discover security issues requiring rapid fixes; coding agents can repair them, but the repairs consume more tokens, and someone remains responsible for the cycle. The next decision is which software to generate, which to buy or reuse, and how to compose both into workflows the customer controls. Flexible vendors that deliver value can still sell software through changing business models.

Broader builders still need specialists

Asked about personal projects, the respondent describes Copilot and an app referred to as Sessions as making unfamiliar codebases easier to inspect, understand, and modify. Broader access expands a person’s scope without implying that everyone should perform the same work.

A project built the previous week was a continuously monitoring chief-of-staff agent. The reported sequence was: 1. Use Work IQ for organizational context. 2. Build a long-running Foundry agent. 3. Add persistent memory through a backend service. 4. Publish the agent to Teams. The respondent says both construction and publication succeeded, connecting the use case to the demonstrated Scout capability for this kind of monitoring.

An interviewer asks whether engineering might consolidate into a few categories centered on agents, deployment, security, and infrastructure. The response favors experimentation rather than committing to that taxonomy. Two examples show how broader roles and specialist work could coexist: - LinkedIn’s full stack builder discipline brings design, product management, and front-end engineering into broader roles while preserving each person’s specialist strengths. - Excel’s learning infrastructure requires distributed-systems expertise. Building a system in which a reward can be learned is described as a difficult infrastructure problem inside an end-user application team.

Specialists are expected to persist, while generalists gain substantial reach. Building an application could become another knowledge-work activity alongside creating a document or spreadsheet.

Miles: build the system that operates the network

Organizational ambition requires a conception of what work could become. The respondent invokes a distinction attributed to Kevin Scott: making difficult work easier versus making previously impossible work possible. The latter requires identifying outcomes an organization could scarcely have attempted before.

Azure networking provides the example. Microsoft reportedly built more Azure capacity in the preceding fifteen months than in its first fifteen years, though the comparison supplies no capacity unit or numerical breakdown. The networking team coordinated more than 500 fiber operators across the wide-area network. Physical cuts and repairs generated communications that needed responses and follow-through.

The team built Miles, an agentic system for networking operations, and reportedly began requesting more tokens instead of more headcount. Its work shifted toward designing the system that performs the operations. The account supplies no measured savings, reliability results, or staffing changes.

The respondent compares this shift to recognizing computer use as knowledge work: calling everyone at a keyboard a typist misses what they accomplish. Likewise, the organizational opportunity lies in using agents to redesign work around meaningful outputs.

Infrastructure must produce visible community benefits

The closing discussion extends the requirement for useful outcomes to data centers. Community acceptance must be earned through tangible local benefits. The respondent sets out several responsibilities for the industry: - Energy affordability and grid capacity: expansion should support communities’ access to affordable energy. - Water stewardship: address replenishment and closed-loop systems. - Local economic benefits: invest in training, construction and ongoing jobs, and the community’s tax base. These are conditions for acceptance, rather than measured outcomes established in the conversation. Community skepticism and hard questions are treated as appropriate.

The token economy should be judged by productivity, broad economic growth and participation, and better health outcomes. The stated goal for the next twelve to eighteen months is that people recognize benefits in their own lives: creating a startup, running a local store more efficiently, or receiving better healthcare. This is an urgent goal, not a forecast of achieved results.

Delegating tasks leaves conceptual learning essential

An interviewer closes by asking why education has shown less impact than they expected, citing OpenEvidence as a healthcare example. That assessment is the question’s premise. The respondent describes a meeting with Alpha School’s founders as useful for rethinking education, while emphasizing the continuing importance of traditional educational work.

The technical example distinguishes learning when and how to apply softmax from simply asking AI to fix a training run. Delegating execution does not remove the need to understand the concepts. Tools that complete a task still leave an educational need to understand the decisions behind it.

What may need rebuilding is the relationship between learning, incentives, credentials, and employment. As access to information and continuous self-education changes, a new university or pedagogy could become a major startup opportunity—if it connects a curriculum to valuable economic opportunities for students.