Launch HN: HyperProbe (YC S26) – Agents that do read-only debugging in prod

Debugging production has always meant one of two miserable options: dig through logs that might not have the line you need, or add a print statement and redeploy, then wait and pray the bug happens again. HyperProbe, launched on Hacker News by founders Shailendra and Karan, wants to kill that cycle entirely by letting coding agents like Cursor and Claude place virtual breakpoints directly into a running service and pull out real variable values the instant something fails.

The pitch is pointed squarely at a problem that's gotten worse as AI writes more of the codebase. Agents are prolific coders but sloppy loggers, so when something breaks, there's often a gap between what actually happened and what the logs captured. Today's fix is an expensive loop: an agent guesses at the root cause using incomplete data, someone adds a log statement, redeploys, and waits. HyperProbe's founders say that loop burns tokens and time, and it's a big reason engineers dread being on-call.

Their solution runs as an SDK inside the service, hooking in-process for Node and Python and attaching as a JVM agent for Java, paired with an MCP server that the coding agent talks to. Tell the agent something like "checkout returns 200 but some orders are failing," and it finds the relevant line, drops a dormant probe there through HyperProbe, and waits. When live traffic actually hits that line, the SDK grabs the local variables across the call stack, redacts sensitive fields like passwords and tokens in-process before anything leaves the container, and streams the result back to the agent for diagnosis. Nothing pauses, nothing writes, and a monitoring layer yanks probes automatically if overhead spikes.

Shailendra and Karan aren't newcomers to this particular flavor of pain. They spent three years building HyperTest, a tool that converted live production traffic into integration tests using OpenTelemetry, which meant they'd already solved a chunk of the hard problem: extracting real runtime state from a running service without breaking it. What they say they learned at HyperTest was less technical and more organizational — testing always got deprioritized the moment prod was on fire, which planted the idea that the real leverage was building tooling for the fire itself, not the fire drills beforehand.

The long-term goal here is bigger than a debugging sidekick. The founders talk about an autonomous on-call agent that takes an alert, probes the system, diagnoses the issue, and fixes it within minutes, no human paged at 3 a.m. required. What's shipping now, supporting Node, Java, and Python, is the first piece of that: give the agent eyes inside the running process instead of asking it to reverse-engineer a failure from whatever scraps of telemetry happened to get logged.