Agent-Native Telemetry: Verifiable State-Delta Evidence for Autonomous Operations
This paper introduces Agent-Native Telemetry, a verifiable evidence architecture using the Agent Telemetry Protocol (ATP) and State-Delta Evidence Ledger to replace verbose human-oriented logs with compact, cryptographically signed state changes, thereby drastically reducing context costs and query overhead while ensuring provenance and security for autonomous AI agents.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the vast, humming infrastructure of the modern internet, where countless digital services run on invisible networks of code, a quiet crisis has been building. For decades, the way these systems reported their health was designed for human eyes. When a server encountered a problem, it would write a line of text, often repeating the same words and labels over and over, creating a massive, scrolling diary of events. This approach worked well when engineers were the ones reading the logs, scanning for keywords to find a broken link or a crashed program. But the landscape is shifting. Today, the primary consumers of this data are no longer just people; they are autonomous artificial intelligence agents, software programs capable of reasoning, diagnosing, and fixing problems without human help. Feeding these intelligent machines the same old, wordy text logs is like asking a supercomputer to read a novel to find a single number; it wastes the machine's limited attention span on unnecessary words and leaves it vulnerable to confusion.
A team of researchers has proposed a fundamental change to how machines talk to machines, introducing a new way of recording system events that strips away the human prose and leaves only the essential facts. They call this "agent-native telemetry," a system built not on sentences, but on verified changes in state. Instead of writing a paragraph about a server slowing down, the system records a precise, cryptographically sealed note that says, "The speed changed from X to Y at this exact moment." This new method, tested on complex networks of software, proves that by removing the fluff and adding a layer of mathematical proof, autonomous agents can understand their environment faster, cheaper, and with far greater accuracy than ever before.
The problem with the current standard is that it is bloated. In a typical cloud environment, systems generate terabytes of log data every day, much of it consisting of static labels like "timestamp" or "service name" that never change, repeated millions of times. When an artificial intelligence agent tries to analyze this data, it must spend its most valuable resource—its context window, the amount of information it can hold in its mind at once—parsing these repetitive words. This leaves less room for the actual reasoning needed to solve a problem. Furthermore, because these logs are just text, there is no way for an agent to be mathematically certain that a specific error did not happen; it might just be that the log was lost or hidden. This uncertainty forces the agent to guess, and in the high-stakes world of cloud operations, guessing can lead to catastrophic failures.
To solve this, the researchers designed a new protocol called the Agent Telemetry Protocol, or ATP. Imagine a system where, instead of writing a diary entry, a machine simply stamps a numbered, sealed envelope containing only the facts that changed. The core of this system is a ledger, a permanent record that organizes operational facts into four specific types of evidence: transitions (when something changes state), observations (a snapshot of a measurement), relations (how things are connected), and checkpoints (a summary of the current situation). Crucially, any messy, unstructured text, such as a long error message or a user's web request, is locked away in a separate, sealed container. The main record only contains a tiny, verified reference to that container. This ensures that the agent's reasoning is never hijacked by hidden instructions buried inside a long string of text.
The researchers built a complete working prototype of this system and tested it against two real-world, complex software environments: a hotel reservation system and an online astronomy shop. They pitted their new system against the current industry standard, which relies on verbose text logs and structured JSON data. The results were striking. By switching to this new, compact format, the amount of data that had to be sent over the network dropped by 96.4 percent. This reduction translated directly into massive cost savings; the researchers calculated that for a large enterprise, the daily cost of scanning these logs for diagnostic purposes would fall from roughly fifty dollars to less than one cent.
Beyond the savings, the impact on the artificial intelligence agents was profound. When using the new system, the agents required 88.8 percent fewer "tokens"—the basic units of information the AI processes—to understand a situation and find the root cause of a problem. Because the data was so clean and the connections between different parts of the system were explicitly mapped out, the agents could diagnose issues in seconds rather than minutes. In one set of tests, the time it took to identify a problem dropped from over three minutes to less than a minute. The agents also made fewer mistakes, correctly identifying the cause of failures with a higher degree of accuracy than when they were fed the traditional, wordy logs.
Perhaps the most significant finding was the system's ability to prove what did not happen. In the old way of logging, if an agent looked for a specific error and found nothing, it couldn't be sure if the system was healthy or if the log was simply missing. The new system includes a mathematical guarantee of coverage. If the agent asks, "Did this error occur between 10:00 and 10:05?" and the system says no, it can prove with cryptographic certainty that the error did not occur, because the record is complete and unbroken. In tests where the researchers tried to trick the system by hiding data or altering the logs, the new protocol detected every single attempt at tampering. It also successfully blocked all attempts to inject malicious commands into the agents, a common security risk when AI reads unfiltered text.
The researchers demonstrated that this approach works not just in theory, but in practice, across a wide variety of artificial intelligence models. Whether using the most advanced commercial models or smaller, open-source versions, the agents performed better with the new data format. They found the right answers faster, spent less money on computing resources, and were immune to the confusion that usually comes from reading unstructured text. The study concludes that as machines take over the management of our digital infrastructure, the way we talk to them must change. We cannot continue to feed them human-style diaries; we must give them the precise, verified, and compact evidence they need to operate with confidence. This new architecture provides that foundation, turning a flood of noisy text into a clear, trustworthy stream of facts that autonomous systems can rely on to keep the digital world running smoothly.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.