Alex Krentsel

Alex Krentsel, also known as Alexander Krentsel, works on agentic systems that improve their own code and behavior. He is a researcher in Google Research’s Systems Research Group and a UC Berkeley NetSys PhD student advised by Sylvia Ratnasamy. His contributions include the Exoharness project with Ankur Goyal and Martin Casado and coauthored research on how deployment evidence should guide agentic software development.[2][3][6]

Krentsel announced the collective release of Exo on July 16, 2026. He coauthored Exoharness: An Agent Harness Architecture for Fully-Recursive Self-Improvement with Ankur Goyal and Martin Casado. The project distinguishes exoharness, the infrastructure substrate, from exo, a self-improving agent built on it.[2][4]

Exoharness separates two responsibilities commonly combined in an agent harness. Trusted infrastructure owns durable state and resources: event logs, secrets, artifacts, and sandboxes. A replaceable executor handles prompts, model calls, tool loops, compaction, and approvals. The event log is the source of truth; prompt history is a view the executor derives from it. This preserves a record when executor code changes, a tool crashes, compaction goes wrong, or an agent edits its own implementation.[4]

The Exo repository describes an agent with visibility into its source code and runtime logs, able to modify prompts, memory, tools, and harness policy, clone itself, and rebuild or restart components. Canonical history lives outside the sandbox filesystem and survives sandbox rewinds, allowing the agent to reconstruct earlier experiments. The project characterizes this evolution as “(mostly) safely”; it remains early software, and the exoharness public API is unstable.[5][4]

In the September 10, 2026 preprint Reality Is the Final Verifier, Krentsel, Shubham Agarwal, Mert Cemri, Shu Liu, Sidharth Sankhe, Ziming Mao, Matei Zaharia, and Ion Stoica argue that even a formal proof cannot guarantee acceptable deployed behavior. They identify two gaps: requirements only approximate stakeholder intent, and a model of the environment only approximates deployment reality. In their framework, reward hacking exploits omissions in requirements or models, while hallucination widens the gaps by fabricating requirements or environmental assumptions.[6]

The authors propose an assurance-revision loop: when stakeholders reject deployed behavior, evidence from deployment should inform changes to requirements, the environment model, or the evaluator. They identify human judgment and faithful, costly evaluation as the principal bottlenecks. Their position is that these gaps require continuous narrowing; predeployment evaluations remain proxies for acceptable behavior under actual conditions.[6]

Krentsel also contributed to SkyDiscover, a framework for AI-driven scientific and algorithmic discovery, with Shu Liu, Mert Cemri, Shubham Agarwal, and other collaborators. He announced its release on March 2, 2026, followed on September 12 by SkyDiscover-Synthesize for just-in-time system synthesis.[2]

Before these 2026 agent projects, Krentsel’s work addressed network control and cloud-assisted autonomous driving. That earlier research supplies context for his experience with correctness, performance, and the constraints of real-time systems.[2][7]

In October 2024, Krentsel, Peter Schafhalter, Joseph E. Gonzalez, Sylvia Ratnasamy, Scott Shenker, and Ion Stoica argued for offloading parts of time-sensitive autonomous-driving computation to the cloud. Their paper, Managing Bandwidth: The Key to Cloud-Assisted Autonomous Driving, identifies bandwidth allocation as essential to meeting strict latency service-level objectives while maximizing the benefit to the car. This was a research argument for cloud-assisted control, rather than a report of deployed vehicles.[7]

His networking publications include the coauthored A Decentralized SDN Architecture for the WAN at SIGCOMM 2024 and CrossCheck: Input Validation for WAN Control Systems at NSDI 2026. At Google, his work focuses on network control architecture and rigorous reasoning about correctness and performance. He received a B.S. in EECS and a B.A. in Music from UC Berkeley in 2019.[2]