Zack Proser

Zack Proser is an applied AI engineer at WorkOS who builds systems that make software agents more reliable, secure, and practical to supervise. His work spans developer infrastructure, enterprise identity, retrieval-augmented generation, and a central challenge of automated software development: protecting the human attention required to judge whether an agent’s work is actually good.

Proser began his software career at BrightContext and subsequently worked at Cloudmark and Cloudflare. At Cloudflare, he described deploying Workers across production websites and APIs to address security and compliance requirements.

At Gruntwork, he helped build git-xargs, an open-source tool for coordinating changes across GitHub repositories. His 2021 introduction to the project explained how it automates maintenance tasks, commits, and pull requests across distributed codebases.

He subsequently worked at Pinecone, focusing on vector databases, embeddings, and evaluating production retrieval-augmented generation systems. At WorkOS, he moved from developer education into applied AI, developing internal workflows and writing about authentication for MCP servers and a documentation server that gives agents access to WorkOS resources. He also coauthored Identity Management for Agentic AI, addressing authentication, authorization, and security for software agents.

• Human attention is the operational bottleneck. Proser argues that agent throughput means little when engineers cannot adequately review the resulting work. His approach to sustainable agent supervision combines Slack and Linear integrations, voice interaction, remote oversight, and deliberate breaks. He cautions engineers to delegate only work they understand well enough to evaluate.

• Closed-loop verification makes autonomy useful. When an internal writing assistant mishandled acronyms including SCIM and SSO, Proser connected a coding agent to the affected Slack workflow and issue-tracking context. The agent implemented a fix, exercised the system, and checked the result. He supplements these feedback loops with verification gates including linting, builds, unit tests, browser checks, and independent review.

• Agent skills should preserve judgment without bloating context. In an open-source workshop developed with Nick Nisi, Proser demonstrates reusable skills built from concise instructions, deterministic scripts, and selectively loaded references. He favors progressive context disclosure and reviewing completed agent sessions to identify repeated friction worth encoding into future workflows.

• Typed workflows make flexible agents dependable. In a TypeScript workshop co-led with Nisi, he demonstrates conversational agents backed by explicit workflow stages, structured outputs, schema validation, and inspectable integrations. His design separates the flexibility of natural-language interaction from the more predictable execution needed in production.

• Economics depend on the completed task. Proser’s coding-agent evaluation harness supports isolated fixtures, grading, telemetry, and spending controls. In a 60-run comparison, he found that the scope of a code change predicted success more clearly than model price within the tested workload. His analysis of inference economics emphasizes cost per completed task, noting that reasoning verbosity, caching, and token mix can overturn advertised pricing advantages—a position he reinforced in a public post about reasoning-token overhead.

His writing on making agent-generated prose publishable applies the same standards to content: factual grounding, voice checks, independent criticism, and safeguards against unsupported claims.