Tejas Kumar

Tejas Kumar is a Berlin-based AI engineer at IBM, the author of Fluent React, and the host of ConTejas Code. His work applies full-stack software engineering to a central problem in enterprise AI: making unpredictable agents reliable enough to handle consequential tasks.

Kumar’s career includes work with Spotify, G2i, Vercel, and Xata, where he was director of developer relations. His 2024 O’Reilly book examines React’s internals, including reconciliation, concurrent rendering, server-side rendering, and React Server Components. His podcast expands his educational work into conversations about web development, AI, and professional life.

He subsequently worked on generative AI at DataStax, which IBM acquired. By 2025, he was applying his frontend experience to AI agents, streaming interfaces, tool calling, and visual workflows. At IBM, his focus includes agent reliability and enterprise retrieval over sensitive information.

• Agent harnesses as reliability infrastructure. Kumar defines an agent harness as the software surrounding a model that anchors its behavior to an observable environment. Its components include tool registries, execution loops, context management, guardrails, traces, and verification. The approach shifts responsibility for dependable execution from prompting alone to the architecture controlling the model.

• Deterministic verification and recovery. In a browser-agent demonstration, Kumar paired GPT-3.5 Turbo with Playwright and Chromium to upvote a Hacker News post. When the agent encountered a login screen and falsely reported success, he added bounded attempts, execution tracing, explicit outcome checks, and a programmatic login handler. The harness handled authentication outside the prompt and verified the completed action, showing how conventional engineering can make an inexpensive model more dependable.

• Privacy-conscious, local-first tooling. His open-source gen-subs generates video subtitles with on-device machine-learning models. Earlier projects include add-gitignore, an interactive project-setup utility, and basically-fullstack-graphql, which explains GraphQL across client and server.

• Dynamically generated harnesses. Kumar has proposed agents that assemble task-specific safeguards before acting, anticipate likely failures, and construct appropriate verification and recovery mechanisms. He presents this as a future direction, not an existing production capability.