$37M+ · AI + Debugging

What they're building

Undo solves one problem: when software fails, developers usually cannot see what it actually did. Most debugging tools show the code as written. They do not capture the code as executed. That gap is where the hardest bugs live — in runtime states that never appear in logs, in race conditions that vanish when you try to reproduce them, in memory corruptions that leave no trace. Undo's product, LiveRecorder, solves this by recording the full execution history of a running Linux program into a single file. Every instruction, every memory access, every system call — captured as a deterministic record. A developer, or an AI coding agent, can replay that recording, step forward and backward through execution, and inspect any state the program was ever in. Non-reproducible failures become 100% reproducible events. The company sells the technology to engineering teams at networking, database, semiconductor and financial services firms. The AI coding agent angle is the new chapter. AI coding assistants have made that distinction matter more. Teams are now shipping far more code than they can fully understand. Undo argues that an AI agent cannot reliably track down why such code fails unless it can see how the software behaved at runtime. Feed the agent that history, the company says, and it can pinpoint the cause of a bug rather than guess at it. The benchmark they publish: on a set of complex bugs, the latest AI models identified the root cause 38% of the time on their own, a figure that rose to 92% when the models were given Undo's runtime recordings. Undo also says the recordings cut token usage on problems the models could already solve.

Why this matters

Palo Alto Networks is a customer. Suresh Sangiah, its senior vice president of engineering, said the costliest bugs in large codebases live in runtime states that logs do not capture. He said Undo often pinpoints the cause autonomously in minutes. "Undo provides the visibility needed to catch and correct errors before they become an operational problem for our customers," he said. The AI coding wave is the market timing argument. GitHub Copilot, Cursor, and their successors are shipping code faster than any previous toolchain in history. That velocity is creating a new class of bug: the AI-generated bug in AI-generated code, in a codebase no human engineer fully comprehends. The traditional approach of print statements, log files, and GDB breakpoints does not scale to that environment. Undo's runtime recording does, because it does not require the developer to know in advance what to look for. Cambridge Judge Business School research from 2013 put the global cost of software bugs at $316 billion per year. That number is larger now, and will be larger still as AI-generated code becomes the norm in enterprise software. Undo has been building the infrastructure layer for this problem since 2005. The $37M is the moment to commercialise at the scale the problem now demands.