AI Aulis

AI NEWS IN ENGLISH · 20 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

41 stories · next issue tomorrow morning · How the paper is made →

← Issue 20 November 2026Paper

Harvey’s patent agent chooses sources to match the information needed and searches within a budget

AI-generated contentWritten by AI

harvey.ai, deepjudge.ai · 6 sources · Sources published: 19 November 2026 · Some source publication dates are unknown

Harvey, a legal AI service, explains in a blog post how its patent agent picks a search source according to the information it needs, switches to a fallback source when necessary and stops at the limit of its research budget. The post includes no measurements of how well the agent works.

In a blog post published on 19 November, Harvey’s Zhonglei Wang and Blake Chizen explain why one search index cannot always answer a patent question. Their example is a question about what claim 8 of a patent covers. A standard search finds the claim as granted. But a lawyer may also need the claim’s original wording in the application and to know whether it is still in force. The lawyer may also need to know how the applicant distinguished the claim from prior art.

This information sits in different places: issued patents, patent office prosecution records and proceedings before the PTAB, the US Patent Trial and Appeal Board. A combined corpus would be costly to build and maintain, the authors write. Even a larger index would not tell the system what information a question requires.

Harvey’s AI agent – a program driven by a language model – works from the question to the information it needs and from there to the available sources. The agent receives the question whole, with patent and claim numbers and dates, rather than broken into keywords.

On every run, the agent receives a routing table of the sources enabled for that run. For each source, the table lists its content, best uses, limitations and fallback. For US patent text, the US patent office, the USPTO, takes priority. If the USPTO is not available in the run, the fallback is Google Patents.

For recurring tasks, the agent has ready-made search paths. To establish a claim’s current status, it moves from the granted claim to any reexamination or PTAB history and finally to the certificate or decision that settles the matter. Harvey has not tried to list every task, because patent questions are, it says, too varied for a fixed decision tree. The agent can build other paths from its goal and its findings.

The agent stops when it can give an answer backed by citations or identify what remains unresolved. It also stops when the run’s research budget runs out. The post does not give the size or unit of the budget. Nor does it include any measurements of the agent’s accuracy, so it is the company’s description of the system’s design, not evidence that it works.

Harvey users can already attach the Patent Research feature as a knowledge source. The same design could later also cover non-patent literature, material from international patent offices, licensed sources and authorised private data, the company says.

Report an error in this story