Alex Volkov is the founder and host of ThursdAI, a weekly AI podcast and newsletter running since 2023, and an AI evangelist at Weights & Biases by CoreWeave. His work focuses on a central problem in AI engineering: how developers retain meaningful oversight as models generate more software and agents operate more independently.
Volkov spent nearly a decade at Fundbox before leaving in late 2022 to build AI products. His earlier Extensionizr simplified browser-extension scaffolding; later projects included a directory of AI image-generation tools, a Colab-based Stable Diffusion setup and an early ChatGPT Telegram bot.
He founded Targum.video, an AI video-translation venture, after experimenting with translating clips shared on social media. His account of the transition traces his departure from Fundbox and the development of Targum using Cloudflare infrastructure.
Production tracing and evaluation. Volkov treats observable execution as a prerequisite for trustworthy AI applications. His work on LLM evaluation emphasizes versioned prompts, retrieval and agent traces, continuous assessment, and application-specific evaluation criteria. He distinguishes programmatic checks, human review and model-based judging, stressing that automated evaluators require their own bias checks. He also favors improving prompts and retrieval before committing to expensive fine-tuning.
Vendor-neutral MCP observability. External Model Context Protocol tools can obscure what an agent actually does across independently operated services. Volkov created the Observable.tools initiative and demonstrated MCP tracing with Benjamin Eckel, advocating OpenTelemetry-compatible instrumentation and trace propagation across tools, programming languages and observability platforms.
The Z/L Continuum. Volkov’s framework for reviewing AI-generated code argues that oversight should match the risk of each change. Authentication, permissions, financial operations and irreversible data warrant close inspection; lower-risk work can rely more heavily on tests, traces, shadow deployments and rollback. He emphasizes independent verification: an agent that writes code, creates its tests and grades its own output has not eliminated review; it has concealed the conflict. As model capabilities improve, human judgment shifts toward defining tasks, designing safeguards and deciding what evidence a change requires.
A chat-with-PDF prototype becomes a production problem when prompt edits break existing features. Tracing, layered evaluation, and application-specific criteria make those failures easier to catch.
MCP tools can hide the work behind an agent’s calls. OpenTelemetry context propagation connects those calls to server execution, while shared conventions make the resulting traces portable.
When adding tools makes an agent harder to observe
Agent-generated code makes implementation cheaper, but review remains scarce. The useful question is which changes need human inspection—and which need stronger systems of proof.
