YC Agents Operate Marriott and Michelin Key Hotel Software With No APIs

The Integration Wall: Why Hotels Are Ripe for Agents

Axelrod Labs, a Y Combinator–backed startup, is running autonomous agents inside the property-management systems of Michelin Key–awarded hotels and Marriott-branded properties across three continents — without a single API integration. Another YC company, Conduit, is deploying the same computer-use approach: agents that navigate legacy PMS screens by sight, clicking and typing like a front-desk clerk, bypassing the integration wall that has protected incumbent hospitality vendors for decades.

Hotel technology spent three decades building walls and calling them foundations. The property management system at the center of most hotel operations was never designed to talk to anything outside its own four walls — it was designed to be the whole building. That architectural choice, made when "integration" meant a fax machine in the back office, has become the largest constraint on modern hospitality software — and the industry is now routing around it.

Legacy PMS platforms remain the operational backbone for over two-thirds of mid-to-large tourism sites, including glamping resorts, heritage hotels, and mixed-use destination campuses, according to field data from EBC Links. Many run on Windows Server 2012–2019 stacks, communicate through ODBC or SOAP interfaces, and demand strict schema alignment for every room status change, rate update, and guest profile sync. In a 2025 global study of 110 chains representing nearly 40,000 properties, 58 percent of hoteliers ranked better integration as their top wish-list item. The HYB Annual Tech Survey confirmed in May 2025 that integration, fragmentation, and missing standards are the most cited problems for operators and vendors alike. Sixty-nine percent name integration as their number-one technology barrier; roughly half cite cybersecurity and integration issues as blockers to PMS implementation.

The technical debt runs deeper than outdated protocols. COIR Consulting finds that many systems running hotels today were built on COBOL, VB6, and TPF, languages with a shrinking pool of practitioners capable of maintaining them, let alone retrofitting them for modern connectivity. Properties accumulate systems sequentially, not by design: a PMS bought five years ago, a channel manager added when OTA volume grew, a POS system chosen independently by an F&B director, a CRM bolted on for a loyalty program. The result is a stack held together by point-to-point connections that break whenever any vendor pushes an update. Ownership of that breakage risk is rarely defined in advance. When it happens, the answer is often "nobody," and the property absorbs the downtime.

Source: EBC Links field-test data, 2024

The economic drag is measurable. Velocity reported in June 2025 that hotels drain up to three-quarters of their IT budgets maintaining outdated systems rather than investing in innovation. A 2025 PMS UX Report found that 52 percent of hotel managers say it takes four months to three years for employees to feel confident using their PMS, and 73 percent of hotel employees are trained in person by a supervisor or colleague, pulling experienced staff off the floor and extending onboarding indefinitely. Hospitality Technology research indicates that 12 percent of restaurant revenue is lost annually to manual data entry errors and disconnected systems; restaurants and hospitality businesses lose five to ten hours per week per location to labor waste from fragmented software. Hotels with full digital stacks achieve RevPAR lifts of roughly 7 percent even in competitive markets, while properties switching from fragmented legacy systems to integrated VoIP save up to £180,000 monthly, based on Velocity's analysis of groups managing 1,000 POTS lines.

"We're essentially trying to teach a rotary phone to send text messages." — COIR Consulting, October 2025

The vendor incentive structure reinforces the stasis. Legacy PMS providers historically monetized their walled gardens through certification fees, installation charges, and ongoing maintenance contracts for third-party connections. Open APIs dismantle that revenue stream by enabling standardized, seamless integrations. NSD Bytes' 2026 analysis states: "If legacy PMS vendors do not immediately open their architectures with robust, bi-directional APIs, they will be completely bypassed and disrupted by agile, AI-first product ecosystems."

That bypass is no longer theoretical. AI agents now require structured data using schema.org markup, real-time APIs with latency under 100 to 200 milliseconds, and protocols like Model Context Protocol for seamless interaction. Hotels that cannot meet those standards risk becoming invisible to AI systems, which prioritize clean, actionable data. Without open APIs, real-time AI agents simply cannot function; they need uninterrupted, sub-second access to housekeeping status, real-time folio balances, and guest preference histories to orchestrate check-in experiences. The lag inherent in API-only integrations (15–60 minutes) and the data drift between systems create a window where overbooking and rate mismatches thrive. For a property managing rate parity across dozens of OTA channels, that window is exactly where revenue leaks.

Two YC Startups Betting on Browser Agents

Y Combinator's last two batches produced two startups taking the same bet: the fastest way into a hotel's property-management system is not an API — it's the same graphical interface the front desk uses every day.

Axelrod Labs (YC S26) calls itself "autonomous software for hotels that takes over all front-of-house and back-of-house tasks with zero integrations." Its agents operate inside the browser, navigating legacy PMS screens by sight, clicking, typing, and reading tables the way a human would. The company says it is live at Michelin Key–awarded properties and Marriott-branded hotels across three continents. TheAgentTimes reported a "four continents" figure when YC announced the backing. Founderland.ai and RuntimeWire both describe the approach as computer-vision–driven browser agents that handle front- and back-of-house work inside existing systems without API keys, middleware, or vendor cooperation.

Conduit positions itself as a platform rather than a single agent. Its "harness purpose built for hospitality" lets operators create, test, and run multiple agents: chat and voice agents that answer guest questions, update reservations through connected tools, and hand off to staff; internal agents that automate operations across the PMS and other hotel software. Conduit's site describes the result as "on-brand agents [that] handle the full guest journey, converting inquiries into bookings and driving upsell revenue, then runs internal operations and surfaces insights for leadership." The distinction matters: where Axelrod sells a managed autonomy layer, Conduit sells the scaffolding for operators to build their own.

All three share a deployment model that would have been impractical two years ago. Large multimodal models can now read a legacy PMS screen, infer the workflow, and execute it reliably enough for production use. That capability turns the industry's oldest constraint, fragmented, closed, API-poor software — into an opening. The incumbents' moat was the integration cost; the new entrants' wedge is the integration bypass.

The pattern is visible in the customer language. Axelrod emphasizes "zero integrations" and "trusted by Michelin Key awarded and Marriott hotels." Conduit emphasizes "connected tools" and "automate operations across your PMS and tools," phrasing that admits the PMS stays put while the agent works around it. Neither pitches rip-and-replace. They pitch a layer that sits on top, learns the clicks, and gradually owns the workflow.

YC's concentration of bets here is not accidental. The batch model surfaces parallel experiments; when multiple companies in adjacent batches converge on computer-use agents for hospitality, the signal is structural, not anecdotal.

Incumbents Are Not Standing Still

The agentic wave has forced every major hospitality-technology vendor to articulate a response — or risk watching the functions they monetized for decades get absorbed by a layer of software that doesn't ask for an API key. Private equity capital is flowing toward agentic AI that handles pricing, guest communications, and group proposals autonomously, and legacy vendors built on fragmented systems are losing ground fast, said Skift's Vivek Bhogaraju, who tracks where capital moves. The bottleneck, operators and investors agree, is data quality: agentic systems only work when the underlying records are clean enough to trust.

Oracle Hospitality, with more than 45 years in the sector, is the most visible incumbent mounting a structured defense. In July 2026, Loews Hotels & Co expanded its Oracle footprint, adding OPERA Cloud Central services to enhance distribution and guest recognition across its U.S. properties. The rollout includes OPERA Cloud Distribution for rate and ancillary optimization across direct and indirect channels, OPERA Cloud Loyalty to drive direct bookings through personalized incentives, and a unified platform spanning property management, distribution, service interactions, loyalty, and sales. Oracle's pitch: a centralized data hub, cloud-based, mobile-enabled, with open APIs — lets properties "get a comprehensive view of its data and business to accelerate innovation that benefits its staff, guests, and bottom line," as Loews put it in the announcement. The strategy amounts to a consolidation play: make the PMS the single source of truth so that any agent, whether Oracle's own or a third party's, has reliable data to act on.

Amadeus faces the same structural pressure; its hotel division sits atop a distribution and itinerary stack that agents could bypass if they can navigate a PMS directly. Cloudbeds and Mews built their reputations on modern, API-first architectures, precisely the integration layer that computer-use agents render optional. All three have signaled that autonomous workflows, including rate optimization, guest messaging, and upsell execution, are moving from roadmap to release in 2026.

The common thread is a shift from "open platform" rhetoric to "controlled agent environment" execution. Incumbents are racing to become the orchestration layer themselves, betting that hoteliers will prefer a vetted, supported agent inside their existing contract over a YC-backed startup that clicks through the GUI. Whether that bet pays off depends on data hygiene: the vendors with the cleanest, most unified guest profiles will field the most reliable agents. The rest are patching legacy schemas while the agents learn to click faster.

Computer Use as a Bridge Layer: The Four-Part Loop

The automation playbook for the past decade assumed an API. Integrate the systems, orchestrate the calls, put a workflow engine in the middle. That advice fails precisely where the money is, because the systems holding the most valuable process knowledge in most enterprises are the ones nobody can integrate with. Robotic process automation was built for exactly this gap. The global RPA market generated roughly $3.8 billion in revenue in 2024 on healthy year-over-year growth, and enterprises are still buying it. What everyone who has operated RPA at scale knows: the license cost was never the real cost. The real cost was maintenance. A bot records a click path against specific selectors and coordinates; the vendor ships a UI refresh on a Tuesday; forty automations break at once.

Computer-use agents are the same premise with different machinery. Instead of recording a selector path, a multimodal model takes a screenshot, decides where to click, moves the mouse, types, takes another screenshot, and repeats. The key difference is how they identify what to interact with. Traditional RPA tools use coded selectors, such as XPaths, CSS selectors, pixel coordinates, and element IDs, that reference specific UI elements programmatically. When the application changes its layout, those selectors break. AI agents use vision models instead. A traditional RPA script clicks at coordinate (420, 280) because that's where "Submit" was when the script was recorded. A computer-use agent clicks the "Submit" button wherever it is because it understands what it's looking for. When a UI changes, the vision model still identifies the same elements by their visual context. A "Submit Order" button that moves from bottom-left to bottom-right of a form is still recognized as the submit action.

This visual approach makes AI agents capable of automating applications that traditional RPA cannot touch: legacy Windows desktop apps, Citrix environments, Java thick clients, mainframe terminal emulators, and proprietary industry software with no documented interface. Desktop agents run on virtual machines or dedicated workstations, the same environments where your legacy applications already run. The rule is simple: if there is an API, use it. API automation is faster, more reliable, and cheaper to maintain. Desktop automation exists for the applications where API automation is not an option.

The architecture that makes this practical is a four-part loop: vision model, action planning, action execution, and verification. OpenAI and Anthropic both ship official computer-use tooling, and the benchmark ecosystem is getting more serious with efforts like WebArena-Verified and OSWorld-Verified. OSWorld matters because it is not a sandbox. It is 369 tasks executed on real operating systems against real applications. When the paper was published in April 2024, the human baseline was 72 percent and the best agent scored 12 percent. Two years later the public OSWorld leaderboard has top entries in the mid-80s, comfortably above that human baseline. But METR's time-horizon research measures capability in the length of task a model can complete rather than in a single accuracy score, and the falloff with duration is steep. In METR's March 2025 measurement, current models have almost 100 percent success on tasks taking humans less than 4 minutes, but succeed less than 10 percent of the time on tasks taking more than 4 hours.

Reliability compounds brutally. Ninety-nine percent per-step reliability across a 40-step process yields roughly 67 percent end-to-end success. At 98 percent per step, it drops to 45 percent. When an RPA bot broke, it broke loudly and stopped. An agent driving the same screen does not halt. It reasons about what it sees and proceeds. The process completes. The log says success. The damage is discovered by a human three weeks later during reconciliation. Indirect prompt injection is structural, not a product defect, and has no complete fix today. Documented 2026 attacks include zero-click hijacking and credential extraction from an authenticated session. A GUI agent inherits the operator's full session — simultaneously the value and the attack surface.

Cost structure follows a different curve than API calls. The WebMall benchmark puts a GUI agent task at 7 to 12 cents open-weight and 34 to 56 cents with frontier models. Input consumption runs 120,000 to 320,000 tokens per task with two- to six-minute runtimes. Fifty thousand invoice lines a month at thirty cents each is $15,000 a month, $180,000 a year, before counting the cost of failures and the humans checking for them. High-volume stable processes favor integration; low-volume high-variance processes favor agents. CUAs are significantly slower than API integrations. A CUA must wait for the screen to render, capture screenshots, and process visual data, which takes seconds per step. If an API is available, always prioritize it over visual automation.

The deployment pattern that works in production is disciplined. Minimum viable deployment: isolated VM, scoped service account, credentials held by the harness, egress allowlist. Anthropic's own computer use documentation recommends a dedicated virtual machine or container with minimal privileges, allowlisted network access, keeping sensitive credentials out of the model's reach entirely, human confirmation for consequential or consent-bearing actions, and review of action logs. Never run a CUA on your local workstation or production server. Always run visual agents in isolated, containerized virtual environments. Before allowing a CUA to execute writes, test it on read-only tasks. Every screenshot, click coordinate, and typed string must be logged.

The implementation pattern practitioners trust is simple: use the agent for navigation and recovery, use deterministic code for verification, and require humans for risky writes. GUI agents belong where variance lives; deterministic code belongs where consequences live. Start read-only, and stay there longer than feels necessary. Record and promote: let the agent run the process, capture the full trace of what it did, and once a stable path emerges, promote that path into scripted, deterministic code. Everything that writes needs tiering by blast radius and reversibility. Instrument cost per successfully completed task, not cost per token or per run.

This compatibility shim, computer use for systems that predate the API economy — is why the hotel story matters beyond hospitality. The same legacy density that makes property-management systems resistant to integration defines defense logistics, aerospace maintenance tracking, and energy grid operations.

Why This Pattern Moves Beyond Hotels

The hotel front desk and the defense program office share a structural problem: mission-critical software that cannot be replaced, rewritten, or API-wrapped on any acceptable timeline. In hospitality, that software is a property-management system from the 1990s; in aerospace and defense, it is the ERP, PLM, and MES stacks that run aircraft production, satellite integration, and shipbuilding. Both sectors face fixed-price economics that punish coordination overhead. Both are regulated enough that "move fast and break things" is a non-starter. And both are now seeing the same workaround deployed: computer-use agents that operate the legacy GUI directly, no integration project required.

The fiscal year 2026 defense budget carved out a dedicated line for autonomy and artificial intelligence — $13.4 billion within a trillion-dollar top line, a figure that alone exceeds NASA's entire annual budget. NATO's collective defense expenditure crossed $1.4 trillion in 2025, with digital command, control, and autonomous capabilities absorbing a growing share. Yet two-thirds of A&D AI initiatives remain stuck as proofs of concept, and a similar share of firms report difficulty aligning AI efforts with firm-wide strategy, per BCG's 2026 survey. The gap between appropriation and production is where the hotel pattern reappears.

Fixed-price development contracts now represent 62 percent of new DoD awards, up from 34 percent in 2021. Under cost-plus, coordination overhead was reimbursable; under fixed-price, every dollar spent on status reviews and schedule arbitrage comes directly out of margin. The Department of Defense spent $9.6 billion on program management office overhead across its top twenty acquisition programs in FY2025. Nearly half funded coordination labor: status reviews, risk adjudication, resource reallocation across subcontractors. Three of the five largest primes are running production pilots that eliminate those coordination roles entirely, replacing them with agentic orchestration layers that autonomously negotiate tasking, budget, and timeline across federated engineering teams.

Lockheed Martin's Skunk Works disclosed in a February 2026 investor call that its Next Generation Air Dominance engineering stream reduced milestone variance by 41 percent after deploying a "resource arbitration agent" across six Tier 1 suppliers. Northrop Grumman's B-21 sustainment contract embeds autonomous logistics agents with authority to re-route parts, re-sequence maintenance events, and dynamically re-price supplier commitments within pre-negotiated bands — no human program manager approval required for decisions under $500,000. RTX reported a 27 percent year-over-year reduction in program management labor costs in its Missiles and Defense segment, attributing the decline to "autonomous decision support infrastructure and digitized governance frameworks." Boeing's MQ-28 Ghost Bat program employs a zero-trust mesh architecture that lets autonomous mission planning agents request and receive airspace deconfliction, tanker scheduling, and targeting updates from Australian Defence Force systems. The first operational squadron achieved initial operating capability in March 2026.

These agents do not sit on top of clean APIs. They operate inside the same legacy environments the hotel agents navigate: SAP, Oracle, Siemens Teamcenter, PTC Windchill, Dassault 3DEXPERIENCE, systems where API coverage is partial, authentication is fragmented, and change control boards measure cycles in quarters. The Department of Defense's CMMC 2.0, which reached full operational capability in January 2026, mandates zero-trust principles for all contractors handling Controlled Unclassified Information. That regulatory forcing function has accelerated zero-trust fabrics that, as a side effect, provide the security substrate for autonomous agents to operate across organizational boundaries. Agents can now negotiate tasking with a coalition partner's logistics system, verify cryptographic credentials via distributed ledger-based identity registries, execute the transaction, and log an immutable audit trail — all without human intermediation.

BCG estimates that using the right AI for the right modernization task can reduce costs by 25 to 35 percent and accelerate implementation by 30 to 40 percent.

The screens are different. The clicks are the same. The vendors who built those screens never imagined software that could read them — and they still don't have to approve it.

Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe and Waymo, and the people building the field.