Offensively or defensively — that is the question to ask of any patent family, and Anthropic's computer-use pair answers "both, with the emphasis still on defensive." The family is two granted records: US12387036B1, "Multimodal agent for efficient image-text interface automation" (August 12, 2025), and US12619815B2, "Magnitude invariant multimodal agent for efficient image-text interface automation" (May 5, 2026). Same core technology, overlapping inventors, less than a year apart. That cadence is itself a strategic signal: the company is actively building a fence, not filing once and moving on.
The licensing posture, inferred from the structure, reads as defensive-first. A foundational agent grant plus a robustness-refinement continuation is the classic shape of a portfolio built to secure freedom to operate — to ensure the company can ship its own computer-use product without being blocked — while accumulating claims that could be asserted if a competitor's agent reads onto them. The "magnitude invariant" refinement in the second grant narrows toward a specific, hard-won robustness mechanism, which is more consistent with protecting a real implementation than with a broad offensive land-grab.
What distinguishes held from asserted matters here, and this brief is careful about it. Anthropic holds these grants; there is no public indication in the patent record that it has asserted them against anyone. Holding a granted patent is a competitive instrument in waiting — it shapes the field by its existence, deterring close copying and giving the holder a bargaining chip — but it is not litigation, and nothing here alleges infringement by any competitor. The strategic read is about posture, not a pending case.
The competitive context sharpens the read. Computer-use and agentic automation are among the most contested frontiers in AI right now, with multiple labs shipping agents that drive interfaces. In that environment, granted IP on the perception-action loop is a real differentiator — not because it stops anyone today, but because it stakes a claim on the mechanism while the category is still forming. The lab that holds the foundational agent grants enters any future licensing or cross-licensing conversation from a stronger position.
The measured conclusion, which this desk prefers to prediction: Anthropic has built a small but coherent granted family around computer-use, structured to protect its own product and to hold optionality against others. Whether that optionality ever becomes an assertion is unknowable from the record — predicting case outcomes is not what we do. What the documents support is narrower and firmer: the family exists, it is granted, and it treats the agent's perception-action loop as core IP worth fencing.