Why Login Metrics Break When AI Agents Do the Work

Photo by Tyler on Unsplash

The Common Belief: Daily Active Users Measures Whether Anyone Cares

What if the single number your product dashboard is built around stops counting the customers who matter most?

That is the uncomfortable question behind remarks from Lofty AI CEO and co-founder Andrew Wild on agents, the Model Context Protocol, and what he frames as the end of login metrics — coverage surfaced via Google News, originally reported by The Tech Buzz. Our read: the claim is directionally right and operationally premature, and the gap between those two things is where engineering teams are about to lose a lot of time.

Start with the conventional wisdom. For roughly two decades, SaaS product teams have treated daily active users (DAU — the count of unique accounts that log in and do something on a given day) and session frequency as the vital signs of a product. Investors ask for them. Board decks lead with them. Growth teams are compensated on them. The entire apparatus assumes a human sits down, authenticates, and clicks.

Lofty AI is an interesting place for that assumption to crack. The platform offers fractional U.S. real estate — blockchain-tokenized shares of rental property, with daily rental income distributed to shareholders, and per its own materials, investments starting as low as $50. A $50 minimum is not a product for someone who logs in every morning to admire a dashboard. It is a product for someone who wants exposure and then wants to stop thinking about it. The human engagement loop was always thin here. An agent doing the browsing, screening, and monitoring does not break the metric so much as finally expose that the metric was measuring the wrong thing.

What Is Actually Verifiable, and What Is Interpretation

Separating the two matters, because the sourcing here is uneven.

Verifiable: Anthropic published the Model Context Protocol on November 25, 2024, open-sourcing it as a standard for connecting AI assistants to the systems where data lives — content repositories, business tools, development environments. Anthropic's own announcement positions MCP as a universal standard rather than a product feature, and the specification lives on GitHub. Anthropic's Claude desktop app connects to local and remote data sources through it. As of November 4, 2026, that two-year-old protocol has become the connective tissue most agent tooling assumes.

Also verifiable, and more telling than any quote: Lofty's own website states that machine-readable context is published at /llms.txt, /llms-full.txt, /.well-known/reasoning.json, and /.well-known/ai-manifest.json. That is not a prediction about agents. That is a company shipping a parse surface for them. A firm that publishes a reasoning manifest has already decided its next cohort of visitors will not be using a browser.

Interpretation: the specific framing that login metrics are ending. That is Wild's analytical position, as relayed by The Tech Buzz, not a measured industry statistic. The research behind this piece contains no figure quantifying agent-driven traffic as a share of platform activity — not for Lofty, not for anyone. Anyone citing a percentage here is guessing.

And the divergence across sources is worth naming. Anthropic's material is narrow and technical: MCP is an integration standard, build once, work across platforms. Tech industry analysis, per TechCrunch's reporting on the agent shift, runs further — predicting agents replace traditional app interfaces for many routine business tasks. Lofty's own site is the quietest and most concrete of the three: no manifesto, just four static files. The protocol owner describes plumbing, the press describes a paradigm, and the operator describes file paths. The full picture that no single source gives you is that the plumbing is real, the paradigm is a forecast, and the operators who are moving are moving in small, cheap, reversible ways.

Photo by CDC on Unsplash

The Pattern: An Integration Surface, Not a Login Screen

Strip the vocabulary away and the agentic pattern in play is unglamorous tool-use with persistent context. An agent holds a goal ("find me fractional rental exposure under $500 with daily distributions"), discovers available tools from a server, calls them, reads structured results, and decides what to do next. MCP's contribution is that the developer writes that tool surface once instead of writing a bespoke integration for every AI platform that comes asking.

Here is the arithmetic the surface reporting skips, and it is the actual economic argument for MCP. Consider a platform that wants agent reachability across the major agentic frameworks — Anthropic, OpenAI, Google, and Microsoft are all building their own, per the reporting here. In the pre-standard world, that is four separate integrations: four auth flows, four schema dialects, four sets of docs, four things to break when any vendor ships a version bump. Under MCP, it is one server plus whatever thin adapters each client needs. Even assuming a generous 25% adapter overhead, one-point-two-five units of work replaces four — roughly a 69% reduction in integration surface area. That is the entire business case, and it has nothing to do with intelligence. It is the same reason every backend engineer eventually standardizes on one API gateway instead of four.

The second-order consequence is the one that actually ends login metrics, and it is not that agents log in on your behalf. It is that under MCP, the agent maintains context across sessions, which collapses the number of authentication events to a fraction of the number of useful interactions. Your auth log stops being a proxy for engagement. One token, issued once, can sit behind hundreds of tool calls. DAU does not fall because users left. It falls because the event you were counting stopped correlating with the thing you cared about.

Chart: Illustrative integration workload for reaching four agentic platforms (Anthropic, OpenAI, Google, Microsoft, all building agentic frameworks as of November 4, 2026) — four bespoke integrations versus one MCP server with thin per-client adapters, assuming 25% adapter overhead. The reduction is the business case; the agent's reasoning quality is a separate question.

Implementation: What the Tool Surface Looks Like for a Fractional-Investing Platform

If a platform like Lofty exposed itself to agents, the shape is predictable enough to sketch — and the sketch is where the hype thins out fast.

You would expose read tools first: search_properties with filters for minimum investment, projected yield, and metro; get_property returning the structured facts behind a listing; get_portfolio and get_distributions for an authenticated holder. Those are safe because they are idempotent — calling them twice costs nothing but tokens. Then, at a very different risk tier, write tools: place_order, sell_shares. Most teams shipping MCP servers in 2026 stop before that line, and they are right to.

The order matters more than the vocabulary. Read tools turn your product into something an agent can reason about. Write tools turn your product into something an agent can lose money in. Those deserve different review processes, different rate limits, and probably different quarters.

Note also that Lofty's published llms.txt and ai-manifest.json are a cheaper first step than an MCP server, and worth distinguishing. Static context files are a one-afternoon change that make your platform legible to agents that crawl. An MCP server is an authenticated, stateful service you now operate — with uptime obligations, token budgets, and an audit trail. Platforms that conflate the two end up operating infrastructure they scoped as a content task.

Where It Breaks in Production

A careful skeptic has three objections, and two of them are good.

The first is regulatory, and it is the strongest. Fractional real estate shares are financial instruments. Suitability checks, KYC, and consent are built around a human attesting to things. An agent holding a persistent token and executing place_order raises a question nobody has cleanly answered as of November 4, 2026: who attested? If the login event disappears, the compliance artifact that lived inside it disappears with it. That is not a metrics problem, it is a liability problem, and it is why read-only tool surfaces will dominate for a while in anything touching an investment portfolio.

The second objection is cost, and it is the one demos always hide. Agents do not browse, they poll. A naive screening loop that re-fetches a property catalog on every reasoning step produces context window blowups and a token bill that scales with the agent's indecision, not the user's intent. Worse, read tools that look free invite tool-call loops — an agent that cannot find a satisfying answer will keep calling until something stops it. Caching, pagination, hard call ceilings, and server-side result summarization are not optimizations here. They are the difference between a working integration and an invoice.

The third objection is the metrics panic itself, and this one does not hold up. Losing DAU is not losing measurement. It is a migration. The replacement instruments already exist in any well-instrumented backend: tool invocations per agent session, task completion rate, value per authenticated token, and distribution events — which, for a platform paying daily rental income, is a far better engagement signal than a login ever was. This is the same analytical trap that the SaaS blog flagged around freemium conversion benchmarks: the metric everyone optimizes is frequently the one easiest to collect, not the one closest to the money.

Our analysis: Wild is early rather than wrong. The protocol groundwork has been public since November 25, 2024, the major platforms are all building agentic frameworks, and at least one fractional-investing operator has already published machine-readable context for agents to read. But the end of login metrics will arrive as a slow re-instrumentation driven by finance and compliance teams, not as a dashboard going dark. The teams that fare best will ship a read-only tool surface this quarter, start logging tool-call economics alongside DAU rather than instead of it, and keep write access behind a human until the attestation question has an actual answer.

Bottom line: build the integration surface now because it is cheap and standardized; rebuild the metric slowly because the number you replace DAU with will be the one your board stares at for the next decade.

Frequently Asked Questions

Disclaimer: This article is editorial commentary and educational analysis based on publicly reported facts; it is not financial advice, investment advice, or implementation consulting, and it does not reflect independent product testing. Named platforms are discussed for illustration only. Research based on publicly available sources current as of November 4, 2026.