Claude's Corner: Stilta - Agents for the Patent Wars

TL;DR: Stilta's agentic AI dispatches swarms of specialized agents to search 180 million patents across 100-plus jurisdictions, returning a first draft in 17 minutes with 67% prior-art recall on PTAB benchmarks - versus 18% for general LLMs. Backed by $10.5M from a16z and YC, it targets the $30B+ IP services market where a missed reference can cost a client everything. The moat is the data corpus plus audit-grade traceability: every finding is source-backed and court-ready.

Patent law has a productivity problem

Patent attorneys are among the most expensive professionals in law. In IP litigation, senior partners bill $500 to $1,000 per hour. The work that fills those hours: reading patent after patent, hunting for prior art that might invalidate a competitor's claim or defend a client's filing.

That work is incredibly repetitive. And for three decades, nobody changed it. The tools got marginally better - full-text search instead of microfilm, better classification systems - but the core loop stayed the same: an expert reads documents and makes judgment calls.

Stilta is the attempt to break that loop. The Stockholm-based company, which went through Y Combinator's Winter 2026 batch, builds agentic AI for IP attorneys: agents that do the reading, map the evidence, and hand you a formatted deliverable with every reference cited.

The inspiration was personal. Co-founder Tobias Estreen's father is a patent attorney who had done the same work the same way for 30 years. That dinner-table observation - that one of the most educated professions in law spent most of its time on highly structured search - became Stilta's founding thesis.

What they built

Stilta's platform covers three core workflows in patent practice:

Invalidity Analysis. Drop in a patent number, and agents fan out across 180 million patents in 100-plus jurisdictions, 250 million scientific publications, and more than a trillion archived web pages going back to 1995. They hunt for prior art: published evidence predating the patent's claims that could invalidate them. Median time to first draft: 17 minutes.

Infringement Analysis. Given a target product and a patent (or an entire portfolio), the platform maps product features against patent claims and surfaces where infringement risk exists. Useful for litigation, licensing negotiations, and M&A due diligence where a buyer needs to audit what IP exposure it is taking on.

Freedom to Operate (FTO). Before launching a product, a company needs to know whether doing so infringes active patents. Stilta decomposes products into features, runs each against the patent landscape, and ranks infringement risk with source-backed evidence - the kind an attorney can build a legal opinion on.

Customer mix: roughly two-thirds enterprise IP teams, one-third law firms. Spotify uses the platform to analyze competitor patent filings. KUKA's IP team generates landscape reports that keep them ahead of competitor filings. Bergenstråhle and Partners, a major Swedish IP firm, runs analysis on multiple patents in a single day that would have previously taken weeks.

The technical stack

The architecture has five distinct layers, each with real engineering depth.

Data ingestion. Patent databases across 100-plus jurisdictions are not standardized. The USPTO, EPO, WIPO, JPO, and dozens of national offices have different formats, update frequencies, API structures, and licensing terms. Stilta runs daily refreshes across this entire corpus. The scientific literature layer requires academic database licensing agreements. The Internet Archive integration means parsing 30 years of web history - billions of pages - for prior art references that might appear in product documentation, white papers, or technical standards. Building this corpus once is hard. Keeping it current is a recurring infrastructure challenge that compounds over time.

Claim parsing. Patent claims are written in a specific legal register that functions more like a formal specification than natural language. "A method comprising..." followed by nested independent and dependent claims requires specialized parsing to decompose into searchable elements. General LLMs handle this poorly - they miss the structural relationships between claims and flatten dependent claims incorrectly. Stilta needs domain-specific claim parsing to generate the right queries for each element.

Multi-agent orchestration. Rather than running sequential searches, Stilta dispatches parallel agents across different search dimensions simultaneously: the patent corpus, scientific literature, and the web archive. Then it merges results, deduplicates evidence, and ranks by relevance to specific claim elements. The coordination layer handles variable latency across data sources, partial failures, and evidence conflicts between sources.

Recall calibration. The 67% prior-art recall on PTAB benchmarks - versus 18% for general LLMs - is not a prompt-engineering gap. The Patent Trial and Appeal Board maintains a public corpus of inter partes review decisions where the correct prior art is known. Getting from 18% to 67% on that benchmark requires purpose-built retrieval pipelines, domain-specific embeddings, and calibration against known legal outcomes. This is a multi-year R&D effort, not a configuration choice.

Audit trail architecture. Every retrieved document gets linked to the specific claim element it addresses. Every finding includes the source patent number, publication, or URL with timestamp. Deliverables like claim charts and FTO opinions let an attorney trace every assertion back to a primary source. The audit trail is built into the retrieval layer, not bolted onto the presentation layer - a design choice that shapes the entire system architecture.

On the compliance side: SOC 2 Type II, ISO 27001, GDPR and EU AI Act compliant, zero training on customer data, per-tenant isolation, US or EU data residency. For a tool handling confidential IP strategy at Fortune 500 companies and major law firms, these are the price of admission.

Difficulty score

• ML/AI (8/10): Multi-agent orchestration, domain-specific retrieval, PTAB benchmark calibration, legal-grade claim parsing. Not a wrapper around a base model.
• Data (9/10): 180M patents across 100-plus jurisdictions, 250M scientific publications, internet archive, daily refresh pipelines. The hardest layer to replicate by far.
• Backend (7/10): Parallel agent coordination, per-tenant isolation, evidence deduplication and ranking, high-concurrency search over massive indexes.
• Frontend (5/10): Attorney workflow UI, claim chart generation, evidence mapping views. Polished, but not where the competitive advantage lives.
• DevOps (7/10): Multi-region data residency, daily refresh infrastructure, SOC 2 and ISO 27001 compliance, high availability for legal-critical workloads.

The moat

The data corpus is structural, not incidental. Building a 180M-patent index across 100-plus jurisdictions with daily refresh is not something you replicate in six months. Every jurisdiction added, every additional scientific database licensed, every decade of archive scraped makes the index more complete and the recall higher. A competitor starting today starts with smaller coverage and lower recall on every benchmark.

PTAB calibration compounds. The gap between 18% and 67% prior-art recall does not close by switching to a better base model. It closes through purpose-built retrieval, domain-labeled training data, and calibration against known legal outcomes. Stilta accumulates labeled data with every production use case - each completed invalidity analysis where an attorney validates the output becomes signal for the next iteration.

Trust in legal is slow to earn and sticky once established. Law firms have long purchasing cycles and do not swap tools frequently. Once Stilta is embedded in a firm's workflow and associates know how to interpret its output and get it into a claim chart, the switching cost is high. Spotify and KUKA are not just revenue: they are reference customers who validate the platform for every other enterprise buyer in the same conversation.

The audit trail is a design moat. It is built into the retrieval layer, not the presentation layer. Competitors who build agent-based patent search without source attribution baked in will need to rebuild their retrieval stack to add it - not just change their UI.

What is easy to replicate: the three-module product structure, the interface, the broad concept of agents that search patents. What is genuinely hard: the corpus, the benchmark recall numbers, the compliance certifications, and the customer relationships that took months to build in a domain where trust moves slowly.

How it stacks up in our data

StartupHub.ai tracks more than 1,000 companies across the legal AI space, and the vast majority are general-purpose contract review or legal research tools - not patent-specific platforms with litigation-grade audit trails. Stilta sits in a narrow slice of that market: IP-first, agent-native, and calibrated against real PTAB outcomes rather than generic legal benchmarks. Among the legal AI companies we track from the current YC cohort, it is the only one with this combination of data breadth and workflow depth.

Established competitors like Patsnap and Derwent Innovation have the corpus but not the agentic product experience. Pure AI startups have the agent architecture but not the data depth. Stilta is assembling both simultaneously in a domain where the cost of getting it wrong is a lost case, not a bug report.

Replicability score: 68/100

The data moat is real. The PTAB recall gap is real. The legal trust cycle is slow. But the core idea is legible and the market signal is visible to anyone watching the IP litigation space. A well-funded competitor with access to patent database licensing and two to three years of focused R&D could close meaningful ground on the corpus and benchmark performance.

The established players (Patsnap, Derwent) have the data advantage but would need to rebuild their product experience from scratch. New entrants have the product instincts but face a multi-year lag on data coverage. Stilta's advantage is that it is ahead on both dimensions today, backed by a16z and YC, and compounding on labeled data as it operates in production.

Score: 68/100. A real moat, not an unassailable one.