Roadmap
What ships today
This release is the evaluation harness and the tool-definition-drift detector. It is the reproducible benchmark that replays labeled agent traces through candidate signals and publishes their operating curves, plus the one detector precise enough to run in production. The signals page records which detectors clear the bar and which do not.
What is deliberately deferred
The mission is visibility — see what your AI agents access, call, and send — and the full expression of that is a runtime library that instruments a live agent, not just a benchmark that scores recorded traces. That layer is deferred on purpose. It would include:
- Framework adapters — hooks into agent frameworks such as the Vercel AI SDK and MCP, so a running agent's tool calls become checkable events.
- OpenTelemetry emission — emitting those events as standard telemetry, so an agent's behavior can be inspected with existing observability tooling.
- A timeline UI — a view of what a session read, called, and sent, so a claim about an agent's behavior can be inspected rather than trusted.
Why it waits
Instrumenting a live agent is only worth the cost once there is a detector worth instrumenting it with. Building the runtime layer around a signal that cannot carry a precision claim would ship visibility that produces false alerts, which is worse than no alert.
The measurements here are what decide the order. Two of the three signals do not clear the bar today, and the one that does is measured on synthetic data with the caveat stated. The honest position is to publish the harness and the results first, and build the instrumentation layer when a detector earns it.
Ship the measurement, let it say plainly which detectors work, and defer the runtime layer until one of them is good enough to justify it. The benchmark is how that judgement stays honest.
Follow the work at github.com/Vedant1202/ashborn.