Cole Murray

Cole Murray is an AI/ML engineer, consultant, and creator of OpenInspect, an open-source background coding-agent system inspired by Ramp’s Inspect. His work moves coding sessions into cloud development environments where teams can collaborate and review the resulting pull requests. He founded CM Engineering, through which he helps B2B SaaS teams build production AI systems, strengthen LLM security, and improve engineering workflows.[1][3][4][5]

In his May 28, 2026 Latent Space interview, Murray traced OpenInspect’s origins partly to friction he observed among consulting clients. A product manager could launch a coding session through Slack, but engineering could not see or join it because it belonged to the invoking user. Passing the work along required copying responses rather than handing over a shared working context.[1]

Murray already had related experimental infrastructure when Ramp published its Inspect write-up. He described comparing Claude and GPT as they worked from that account, using his existing understanding of the system to verify the results. He announced OpenInspect on January 29, 2026, presenting a self-hostable foundation with multiplayer sessions and Slack and web clients that companies could customize.[1][11]

OpenInspect keeps session state in its control plane rather than in an individual client. A task can begin in Slack, be watched in the web interface, and finish as a GitHub pull request; colleagues can join the same session and refine the work. Git commits credit the prompting user as author, with Open-Inspect recorded as committer.[6][12]

Preparing reproducible development environments is a substantial part of Murray’s consulting. He described teams whose setup depended on asking a colleague for secrets—an arrangement an autonomous agent could not reproduce. OpenInspect provides provisioning and startup scripts, prebuilt images, and, on supported backends, snapshots retaining dependencies, built artifacts, and workspace state. These let subsequent work reuse an environment instead of repeating its setup. In his project account, he emphasizes engineering approval: “The goal is not to skip review. The goal is to move more work into a reviewable form.”[1][6][12][3]

Murray sees company integration as the larger step beyond deploying an agent: connecting it to logs, read-only database access, and internal knowledge. In May, he described SRE triage as a common client use case, with agents gathering context before engineers respond and sometimes proposing a pull request. OpenInspect supported Sentry and generic-webhook triggers for this work. He also reported PMs requesting fixes through Slack and support teams handing engineering an investigation’s context; first response did not necessarily mean autonomous incident resolution.[1]

OpenInspect separates a control plane from the coding environment. Cloudflare Workers, Durable Objects, and D1 coordinate session state, streams, integrations, stored secrets, and sandbox lifecycle; the data plane runs the development environment and coding harness. In May, Murray credited an unnamed contributor with adding Daytona support alongside Modal, while E2B remained on the roadmap. By September 18, the project documentation also described E2B support.[6][1][12]

Murray’s architectural preference differs from the implementation. In the May interview, he favored running the agent’s “brain” outside the sandbox, issuing tool calls into it. He argued that this improves security separation but makes state management more complex; keeping the harness inside localizes its state. He contemplated moving OpenInspect’s harness outside, while the September documentation still described OpenCode running inside the sandbox.[1][12]

OpenInspect’s security model assumes trusted members of one organization. A shared GitHub App supplies repository access, and the system does not validate each user’s repository permission when creating a session. Scoped secrets are encrypted in storage and injected into sandboxes at startup. This makes the organization and the App’s repository scope material trust boundaries.[3][12][6]

Murray’s open-source decision reflected his expectation that background agents would become critical company infrastructure requiring customization. In May 2026, he expressed reluctance to turn OpenInspect into a service competing at a hypothetical $20 per seat. He considered the intermediary position between model and sandbox providers difficult to monetize, with integrations and infrastructure among the remaining things to sell. The project site offered tailored deployment by September, alongside the open-source, self-hostable system.[1][6][14]

In his July 29, 2026 build-versus-buy essay, Murray argued that the valuable asset is the ability to change the system when agents fail. He identified environment capture and company integrations as the difficult work, including access controls, attribution, and deployment topology. Hosting a vendor’s execution workers, he argued, does not give a team control over its planning and orchestration. His recommendation has an operational condition: “if nobody is going to own the loop, buy.” He also said adopting another platform beats prolonged planning for teams not yet running agents.[7]

Murray advocates continued human inspection of agent-written code. His warning that “your code base regresses to your worst engineer” describes how unaudited patterns become examples that AI repeats: long conditional chains or duplicated date-formatting helpers can spread through a repository. He recommends scheduled cleanup and described implementing getattr lint rules with clients. He recommended Semgrep or other tools for excessive compatibility scaffolding and linting for weak typing, interpreting some of these patterns as attempts by models to avoid failure.[1]

He also described restraint around agent memory. In May 2026, Murray said difficult retrieval problems had made him hesitant to add general memory to OpenInspect. Skills and updates to Claude.md served as partial substitutes among his clients. He considered integrated AI review important as well: OpenInspect had a configurable GitHub reviewer, but bot follow-up on comments or merge conflicts then required a request or tag.[1]

Before his background-agent work, Murray worked on recommendation systems, manufacturing AI, mobile software, and wearable sensing. Those earlier roles provide context for his experience building ML platforms and deploying software.[5]

In his career account, Murray lists Amazon work during 2016–2022, including technical leadership of the Similarities team and work on features such as “Customers Who Bought Also Bought” and “Frequently Bought Together.” He says he built a machine-learning platform that let technical and nontechnical users create custom recommendations and operated a high-availability recommendations service. He also co-founded Empiric, a manufacturing AI/IoT company, where he led product and technology as CTO. He describes winding down that business in 2024 and shifting full-time to CM Engineering.[5]

Murray’s account lists Android engineering at Credit Karma during 2014–2016, including work on a new offers experience and co-development of its mobile analytics system. Earlier, during 2013–2014, he worked on Insight, a smartwatch context-sensing and prediction framework at the University of California, Riverside, under advisor Reza Rewassizadeh. He describes his research contributions as sensor-log generation and retrieval, file storage and I/O, and semantic abstraction.[5]