Agentic AI Inventorship: The Open Question

There is a clean line between two questions that look similar. "Am I still the inventor if AI helped me build it?" is answered. "Who is the inventor when the AI did most of the building on its own?" is not. The first is the subject of our AI inventorship and human conception page. This page is about the second, which is where agentic systems are taking the question.

We will be precise about confidence. Some of what follows is settled law in multiple jurisdictions. Some of it is genuinely unresolved, and where it is, we say so rather than guess. The value of this page is the map, not invented answers to questions the offices themselves have not answered.

One distinction first, because it is the most common category error: this page is about inventorship, not ownership. A company can own patent rights through assignment, employment agreements, and chain of title regardless of who invented. But the application still has to name the natural persons who conceived the claimed invention. Agentic AI does not threaten your ownership. It complicates the question of whether there is a human inventor to name.

This article presents published guidance, case law, and academic and practitioner analysis for educational purposes. It is not legal advice, and it does not predict how any office or court will resolve open questions. Consult a registered patent attorney for filing decisions.

What is settled: major examining jurisdictions reject AI as the named inventor

The settled rule is narrow and firm: the AI system itself cannot be the inventor named on the application. It is not that every human-plus-AI fact pattern has been resolved. The major jurisdictions reached the threshold answer independently through the DABUS litigation, where Dr. Stephen Thaler tried to name an AI system as the sole inventor. None of these is in doubt.

The pattern is consistent: of the jurisdictions to have ruled on whether an AI can be named as the inventor, only South Africa allowed it, and because South Africa does not substantively examine patents, that grant is a formality-system outlier rather than a substantive holding. Australia briefly went the other way at first instance, but the Full Federal Court reversed (Commissioner of Patents v Thaler [2022] FCAFC 62) and the High Court refused special leave, leaving the human-only result in place. China's CNIPA examination practice is generally described the same way, treating inventors as natural persons, though we have not linked a primary citation here and so state it only as the reported position.

The United States layer adds an examination-side reaffirmation. The USPTO's November 28, 2025 revised inventorship guidance (90 Fed. Reg. 54636) states that the traditional conception standard governs every invention regardless of AI involvement. AI is a tool. Conception, the formation in a human mind of a definite and permanent idea of the complete and operative invention, remains the test. We cover that doctrine in depth on the human conception page.

What is open: where does conception live in an autonomous loop?

Here is the gap. The settled rule says a human must conceive the invention. The guidance was written for a world where a human directs an AI tool the way an inventor directs a CAD program or a search engine: the human holds the idea, the tool executes. Agentic systems strain that picture. An agent can decompose a goal, run many steps, evaluate its own intermediate results, change direction, and surface an output the human did not specifically envision and could not have written down in advance.

A goal is not conception. The USPTO's guidance is clear that conception requires a definite and permanent idea of the complete and operative invention, not merely a research objective. That is exactly why agentic workflows are hard: the human may have set the goal, while the agent first produced the specific, claim-relevant solution. When that happens, the doctrinal question has no published answer:

• If the human supplied only a high-level goal ("find a more efficient cache eviction policy") and the agent produced the specific, novel, operative solution, did the human form a "definite and permanent idea of the complete and operative invention," or only of the problem?
• How much human direction, evaluation, or selection is enough to count as conception rather than mere appreciation of the agent's output after the fact?
• If no natural person conceived the claimed invention, there may be no proper inventor to name, even though a human plainly set the process in motion and owns the system. That is not the same as saying every agent-assisted invention is unpatentable. It means the human-conception record becomes the filing question.

The honest status: the current guidance provides no agentic-specific examples, and commentators have flagged exactly this. Analysis of the revised USPTO framework in the Oxford Journal of Intellectual Property Law & Practice notes that unresolved questions remain about applying the conception test to more autonomous contexts. Academic work has begun to treat the autonomy gradient directly; see, for example, the analysis of fluid autonomy in agentic AI and its implications for inventorship. None of this is settled law. It is the frontier being mapped.

Why this is becoming urgent, not academic

DABUS was an edge case: a researcher deliberately testing whether a machine could be named. Agentic inventorship is arriving through ordinary practice instead. Engineering teams are running autonomous coding and research agents that genuinely produce novel approaches with thin human steering. The question moves from a philosophy-of-law hypothetical to a "who do we name on this filing" decision that real teams face now. The DABUS filings asked whether the machine could be named. Agentic workflows more often ask the opposite: whether the human record is strong enough to name a person.

The institutional signals point the same way. The major offices are coordinating on AI rather than acting alone, which usually precedes guidance rather than follows it. Patent offices are also adopting AI operationally for their own work, which is separate from inventorship doctrine: the USPTO has run the ASAP! pilot for AI-assisted pre-examination prior-art search and announced Class ACT, an agentic trademark-classification tool with humans still in the loop. Those show agency experimentation with AI; they do not answer who conceived an invention in an agentic R&D workflow. That combination, fast operational adoption plus open inventorship doctrine, is exactly why teams should build a human-conception record now. The risk is not a retroactive rule. It is that the line gets drawn later and your already-filed record is what gets judged against it.

The practical posture while the law is unsettled

You cannot resolve an open legal question by choosing a workflow. But you can make your record robust to however it resolves, and the direction of that hedge is clear from the one thing that is settled: conception is about what a human formed in their mind. The more autonomous the tool, the more weight the human-conception record has to carry, not less.

In a tool-like workflow, human conception is usually obvious from the work product. In an agentic workflow, it is exactly the thing that can become invisible, because the agent generated the artifact. So the documentation that mattered a little in the assisted case matters a lot here:

• Capture the human's definite-and-permanent ideas as they form. Which specific technical choices did a person settle on, versus accept from the agent without independent judgment? That distinction is the whole question, and it is far easier to record while building than to reconstruct at filing.
• Record direction, evaluation, and selection, not just the prompt. A goal statement alone is weak evidence of conception of the solution. Evidence that a human understood, judged, modified, and chose among the agent's outputs is stronger.
• Treat "the agent did it on its own" as a flag, not a brag. The more genuinely autonomous the path to the claimed invention, the more important it is to get patent counsel involved before filing, because that is precisely where the open question bites.

This is the same discipline the conception note template on our inventorship page is built for, and it is why we keep saying the inventorship problem starts before you file. In agentic workflows that is not a slogan; it is the difference between having a conception record and having an artifact with no human fingerprints on it.

A risk ladder for agentic workflows

Inventorship risk is not binary. It rises with how much of the claim-relevant inventive work the agent did on its own. This ladder is a framing tool, not legal advice, but it maps the workflow to the posture and the response.

The reason the multi-human row points to Pannu v. Iolab and not to the AI: the USPTO's November 2025 guidance restored Pannu to its ordinary role of evaluating joint inventorship among natural persons. An AI is not a person and cannot be a joint inventor, so Pannu is never applied to the tool. See the human conception page for that doctrine in full.

The agentic conception packet

For any invention that came out of an agentic loop, the most useful artifact an engineering team can preserve, under counsel's direction, is a short packet that makes the human conception visible. The goal is not to save every prompt forever. It is to preserve enough evidence to show human conception of the claim-relevant features.

Agentic conception packet

1. Human problem statement
   What technical problem did a person identify?
2. Claim-relevant constraints
   Latency, security, compatibility, cost, data, deployment, reliability.
3. Human technical hypothesis
   Before the agent ran, what solution direction did a person believe would work?
4. Agent role
   Search, code, benchmark, refactor, design, simulate, rank, decide.
5. Human decisions
   Which outputs did a person reject, modify, combine, or select, and why?
6. Alternatives considered
   What other approaches were weighed and rejected?
7. Evidence
   Design doc, issue, PR, commit hash, benchmark, diagram, experiment log, run ID.
8. Claim mapping
   Which claim-relevant features came from human conception, and which first
   appeared in agent output?
9. Contributors
   Which humans made technical decisions, and which merely operated tools?

One caution: prompts, agent outputs, and run logs can carry confidentiality, privilege, and discovery consequences. What and how you retain should be decided with counsel, not by default. Our discovery and evidence page covers that side.

Two traps that catch teams

Both are non-obvious to engineers and founders, and both get worse in agentic workflows.

• Inventorship follows the claims, not the product. The inventor list is set by who conceived the claimed subject matter, and prosecution can move the point of novelty from one feature to another. A contribution that looked incidental during development can become the claim-critical limitation later, which means inventorship should be re-checked after material claim amendments. 35 U.S.C. § 116 allows joint inventors even where they did not work together, contribute equally, or contribute to every claim.
• Naming choices abroad can break U.S. priority. Under the USPTO's guidance, a U.S. application claiming benefit of an earlier application must share at least one natural-person inventor with it. A foreign priority application that names only an AI system can sink the U.S. priority claim. For global teams, the inventorship record is not just a U.S. formality; it can shape filing strategy across jurisdictions.

Inventorship is not the only frontier

Agentic AI stresses more than inventorship. It also pressures obviousness and the prior-art baseline: does advanced AI change what a person of ordinary skill in the art is presumed to know, and when does an AI-generated output become public knowledge that counts as prior art? Those are separate patentability questions, not inventorship questions, but they point the same direction: agentic workflows need earlier patent hygiene. For the eligibility and precedent side of the cluster, see our Section 101 declaration strategy and AI patent legal precedents pages.