Roboto Agents Traces Robot Failures to Responsible Code

Robotics failures are annoying for a specific reason: knowing that a magnetometer glitched at 14:32 tells you almost nothing about why. The real answer usually lives somewhere between hardware, sensor fusion, a firmware config, and the memory of whichever engineer touched that subsystem last. Roboto, the robotics data platform, spent the first two days of its self-styled Agent Week showing off triage agents that read logs and slap labels on problems. Day three is the more interesting move: agents that go read the code too.

The pitch is straightforward. Instead of an agent flagging "communication dropout" and stopping there, it now pulls in the actual repository, tests hypotheses against real flight data, and follows the failure back to the commit that likely caused it. Roboto's example is an estimator config that shipped on a Tuesday and quietly changed how a sensor reading gets fused into the state estimate. A log-only agent would never catch that. One with code access can, and it comes back with a proposed fix and a trail of evidence an engineer can actually check.

The plumbing here is MCP, the open protocol that's become the default way to wire agents into external tools. Point Roboto at a GitLab MCP server and its agents will trace a triage label or a flagged event directly to the responsible code. Or skip Roboto's own agent entirely and hook something like Claude Code straight into Roboto's MCP server, so your own agent queries raw logs and datasets rather than a summarized version of them. That second option matters for teams whose security folks don't love the idea of data leaving their environment.

BRINC, the drone company, is Roboto's proof point. Support engineers there triage issues in plain chat with zero code knowledge, and the same conversation apparently continues into deeper engineering territory without anyone switching tools. Jon Hoff, an autonomy engineer at BRINC, says root-causing a gnarly flight failure used to eat hours or days; now it's minutes, and staff who don't know every subsystem cold can still diagnose problems. That's the kind of claim vendors love to make, but the underlying idea — collapsing a three-team investigation into one thread — is genuinely the harder engineering problem, not the marketing one.

Roboto's Agent Week isn't done; a fourth installment is teased for later the same day, apparently about finding moments in data beyond failures. But this middle chapter is the one that actually says something: robotics companies have had good telemetry for years and mediocre tools for connecting that telemetry to source code. If agents can reliably close that gap, root-cause analysis stops being a multi-day scavenger hunt and starts being a chat message.