
OAuth 2.0 scopes authorize AI agent tool calls at the wrong grain: a scope like calendar.write covers every mutation an agent could ever make, is fixed at request time before the agent knows which specific call it will make, and cannot be narrowed by the holder before delegating to a sub-agent. We argue that capability tokens, i.e. attenuable, call-time-bound authorization primitives with 30-plus years of prior art in the capability-security literature, are the right fit for agent tool-call authorization, and that the existing OAuth 2.1 machinery (Token Exchange, RFC 8693; DPoP, RFC 9449; Transaction Tokens) already provides most of the mechanism needed to implement this without bespoke cryptography. We measure the cost of one concrete implementation across a nine-hop delegation chain and find that attenuation overhead is small and does not grow with chain depth; the dominant per-hop cost is unrelated client authentication.
Organisations hyperlinking with traditional HTTP or file URLs across a set of tools such as office suites or web-based ticketing systems is a common practice that creates an undesirable technical coupling to specific locations, hostnames or applications. By using a new URI-scheme incorporating a tool-specific identifier and the type, a custom protocol handler middleware can resolve these identifiers on-the-fly to traditional URLs. The handler integrates at desktop level with any hyperlinking capable application. Organisations can adopt these scheme links as persistent identifiers for internal use (or employ existing ones). A complementary but more coupled and limited method pre-converts documents to PDF by replacing custom URIs with traditional URLs for read-only distributions without the protocol handler installed. The solution has been deployed it the ISO 13485 quality management system of a medical device with documents residing in web and desktop applications. After a decade 90% of links use the proposed URI.
The Internet's trajectory is rooted in the early 1960s with the introduction of the concept of packet switching. Since that time, the terrestrial Internet has evolved in geophysical scale, speed, scope and application, in some cases by up to six or even seven orders of magnitude. It has escaped the planet to work in Low, Medium and even Geo level satellite orbits. New networking technology has emerged to take the internetworking concept into Solar System scale operation. Artificial intelligence (AI) and AI agent technology have risen from the Internet's infrastructure.
Drone technology has advanced well beyond simple uncrewed flight; modern drones are capable compute engines with reliable wireless connectivity. At the same time, artificial intelligence that reasons and plans can now run on resource-lean devices, allowing drones and nearby edge infrastructure to manage computation locally or offload to data centers, whichever best advances mission objectives. These trends enable aerial edge intelligence, an emerging capability that could enable autonomous fleets of uncrewed aircraft to execute complicated, multi-faceted missions in dynamic conditions. However, realizing aerial edge intelligence requires new computing capabilities for safe, scalable aerial operations. In December 2025, we convened a workshop of 47 experts. This article identifies properties of aerial edge intelligence that existing frameworks miss, shows how they create three critical challenges, and proposes a research agenda for computing on the fly.
UAV-assisted Internet of Robotic Things (IoRT) deployments need to achieve reliable, high-fidelity environmental monitoring in vast, remote areas. Traditional methods, which focus solely on the Age of Information (AoI), fail to account for the diverse volatility of physical processes, leading to suboptimal sampling and missed anomalies. This is because effective monitoring in dynamic environments depends not only on timely data but also on accurate reconstruction of rapidly changing underlying processes. To address this, we propose a process-aware edge intelligence framework that jointly optimizes UAV trajectories and sampling schedules based on the unique dynamics of each sensor, directly minimizing Expected Mean Squared Error (EMSE). Our solution improves both estimation accuracy and anomaly detection compared to baseline approaches, as demonstrated in comprehensive simulations. These results underscore the value of process-oriented, adaptive monitoring and open new directions for resilient, task-driven IoRT systems capable of supporting demanding real-world applications.