Laurie Voss

Laurie Voss is a developer, npm Inc. cofounder and developer relations practitioner whose AI work addresses how to build useful LLM applications and decide what agents should optimize. Writing as seldo, he has advocated grounding models in task-specific information, using ordinary code where possible and keeping human judgment central to application design.[2][3][5][7]

In a March 2026 Arize account, Voss described a team using Claude Code to improve a tool that turned tweets into a newsletter draft. The agent made implementation changes, ran experiments and debugged failures; the humans supplied evaluation criteria and direction. Tests combined code checks for Markdown structure and source URLs with an LLM judge for faithfulness and writing quality, using 100 tweets divided into five batches of twenty.[7]

One problem began upstream of the prompt: the model copied shortened t.co links from the tweet text. Expanding those links before sending the text to the model improved the source-link check from one passing batch to four out of five, according to Voss’s account. Explicit URL-copying instructions and a larger output allowance subsequently brought that check to five passing batches.[7]

The consequential decision concerned the evaluator itself. To satisfy a requirement that every link appear, the agent appended a “Tweet Sources” section containing a wall of URLs. The checks passed, but the humans rejected the result as unhelpful to readers and instructed the agent to remove it. They then replaced mechanical link completeness with semantic tweet coverage: an LLM judge checked whether each tweet’s content was represented, allowing related tweets to be summarized together and trivial posts to be skipped.[7]

Voss reported that the revised suite passed all five batches, with an 80% threshold for the new coverage evaluator. Those results applied to the fixed dataset and a changed evaluation objective. His distinction between implementation and direction was explicit: “The agent wrote all the code, ran all the experiments, and debugged all the failures. The human decided what mattered.”[7]

In his January 2025 essay “What I've learned about writing AI apps so far”, Voss offered a rule of thumb: “LLMs are good at transforming text into less text”. He recommended supplying the material needed for a task—a contract to analyze or a video transcript to summarize—rather than assuming the model’s training contained reliable answers. He explained retrieval-augmented generation as giving the model relevant information and asking it to return that information in condensed form.[5]

Voss also recommended giving models opportunities to inspect and correct their work, while assigning deterministic tasks to ordinary code for speed, cost and reliability. His advice favored human augmentation: “Jobs are more than collections of tasks.” In his account, prioritization, exceptional situations, communication and experience made replacing an entire job a different problem from automating its component tasks. He particularly cautioned against applications intended to replace doctors or lawyers.[5]

In September 2026, “We are all Product Engineers now” extended Voss’s discussion to the future of software work. His ten-year forecast rested on two explicit assumptions: agents would eventually handle the entire software development lifecycle, and demand for software would have no practical upper bound. He distinguished that anticipated capability from present weaknesses in review, maintenance and operations.[6]

Under those assumptions, Voss argued that the bottleneck would move toward discovering and precisely describing what customers want. His example was an order-tracking request: a bakery and a car-parts factory could use the same words while needing different software. Someone still had to uncover the specific need, define what “good” meant and check the result with the customer. He forecast a larger profession increasingly shaped like product engineering, while acknowledging that this was also the part of software development he personally preferred.[6]

Voss’s concern was how people would acquire that judgment. He argued that automating junior code-production work could disrupt the path through which developers learned from senior review and accumulated customer context and taste. He also recognized a loss for programmers who loved the craft of writing code and would not welcome a job centered on customer discovery.[6]

Before these AI essays and evaluation experiments, Voss had a career in web development, company founding and developer communication. In October 2023, he described having started as a web developer 27 years earlier and having founded companies including npm Inc. His work communicating with developers included personal writing, conference talks and community surveys.[4]

Voss joined LlamaIndex as VP of Developer Relations on October 2, 2023. He compared the possibilities of LLMs with the mobile platform shift, emphasizing applications that could not previously have existed. LlamaIndex’s Python and JavaScript framework connected data sources and models for building LLM applications. In his announcement, he named Jerry, Simon and Yi as colleagues he looked forward to working with, noting that Yi would focus more on partnerships.[4]

Voss served as an emcee at AI Engineer World’s Fair 2025, acknowledged in Latent Space’s June 2025 recap. By September 20, 2026, his GitHub profile listed Arize AI, while his personal site described his work as developer relations.[1][3][2]