ena pragma
·8 min read·ai-operations · agent-architecture · ai-verification

Vercel's new agent framework tells a real scaling story. A different 2026 paper argues the opposite bet.

Eve is Vercel's new filesystem-first agent framework. Hypergraph research argues almost the reverse. Both are real, dated, and unconnected in any source we found.

Contents

A Vercel engineer took the stage at AI Engineer this month and told a five-generation origin story about building an internal agent, ending in a new open-source framework called Eve. Separately, a small but active body of 2025-2026 research argues that the exact structure Eve is built around, a flat directory of files, is the wrong shape for a capable agent to compose tools reliably. We checked the talk against ten external sources, checked hypergraphs against ten more, then checked whether the two connect. The honest answer sits in the middle: real, sourced overlap on the underlying problem, and zero evidence the two have ever met.

#What Vercel actually built, five times

The talk, "How We Solved Agent Building," narrates a data-science team's agent in five attempts, in order. First, a single mega-prompt with a database schema pasted in, SQL copied out by hand. Second, a chained pipeline of scoped agents, one for querying, one for planning, one for execution, one for reporting, which improved things before it "started hitting some walls." Third, a single mega-agent that managed its own memory and reflected on its own steps, which tested well internally and then failed once handed to a wider group: "we couldn't have anticipated some of the questions." Fourth, after Claude Opus 4.5 and Claude Code shipped, a rebuild around a minimal filesystem agent, list files, read files, run bash, plus a few Vercel-specific tools. The speaker credits this as "the biggest unlock ever" and says it roughly doubled their internal evaluation score. Fifth, a recurring job that distills repeated query patterns into reusable skill files, because "every new agent run... starts from nothing" without them.

The load-bearing line from the talk, worth quoting directly: "claw code was not giving it a very prescriptive set of tools. It was sort of just letting it go wild and explore emergent behavior." A minimal, general tool surface beat a larger, purpose-built one, once the underlying model got capable enough to use it well. That claim checks out against Vercel's own changelog, which describes the same filesystem convention in the same words: "an agent is just a directory of files," with agent.ts, instructions.md, tools/, skills/, subagents/, channels/, and schedules/ as the shape. The company's own site describes it in a simile, not a slogan: "like Next.js for web apps, but for agents," convention over configuration.

Two things in the talk are worth naming as unaudited rather than confirmed. The claim that a single blog post drove "70% of our vercel.com traffic" the week it ran is Vercel's own internal, self-reported figure; no independent analytics source confirms or denies it, so treat it as color, not as a cited number. And Eve's one named beta customer, Aura, comes with results ("fewer steps, better successes") that are the speaker's own account, not independently verified. The repo itself is real and active: created in June 2026, over five thousand stars at the time we checked, releases landing at roughly daily cadence through the day of the talk itself. Reading the timeline plainly, Eve is the third major implementation of "skills as reusable files," a convention Anthropic introduced in September 2025 and OpenAI adopted with its own catalog two months later. Eve did not invent the idea. It is the first to fold it into a full, opinionated framework.

#The other side of the bet

A hypergraph generalizes an ordinary graph: where a normal edge connects exactly two nodes, a hyperedge can connect any number of them at once. That distinction matters whenever a real relationship is a group fact, not a chain of pairs. Three people co-authoring one paper is one three-way event; forcing it into an ordinary graph as three separate two-way edges loses the fact that it happened once, together.

That is not a new idea, hypergraph neural networks go back to a 2018 paper, but its application to multi-agent LLM systems is recent and active. Five real, checkable papers from 2025 and 2026 make variations on the same argument. One, from Cambridge's Prorok Lab, argues that pairwise message-passing between agents "leads to suboptimal behaviours... particularly in dense environments where group coordination is most critical." Another applies the same critique to how agents communicate rather than how they move. A third, and the one most directly relevant to what Eve is trying to do, represents each tool as a hyperedge running from the inputs it needs to the outputs it produces, so an agent can plan a chain of tool calls by walking that structure directly, "instead of relying on LLMs to infer tool compositions from textual descriptions, which can lead to inefficient exploration and unreliable execution." A fourth applies the same idea to multi-agent memory, arguing that today's dominant memory shapes, vector and graph, "flatten these structures into embeddings or dyadic traces, obscuring events involving agents, tools, documents, errors, and evidence." Nature Communications published a related argument aimed at a different failure mode: reducing LLM hallucination in domain-specific retrieval (benchmarked on a medical dataset), using the same hypergraph-over-flattening logic. This is not a fringe idea circling one lab. It is the same critique, independently arrived at, aimed at coordination, communication, tool use, and memory in turn.

#Where they meet, and where they plainly do not

Here is the real overlap, sourced from both sides. Eve's entire design bet is that a flat directory of files is enough structure for a capable model to compose reliably. The tool-schema hypergraph paper above is a direct, dated, same-year counter-argument to exactly that bet, aimed at exactly the same problem, an agent composing tools reliably. One side says a capable enough model turns flat text into working composition. The other says composition needs explicit structure regardless of how capable the model is. Same target, opposite premise, both published in 2026.

Here is what does not connect, and would be overclaiming to assert otherwise. Nothing we found shows Vercel, the speaker, or Eve's own documentation discussing hypergraphs, citing this research, or planning to adopt it. A direct search combining the two names came back empty. The connection above is a reader's inference sitting two sourced facts next to each other, not a fact either side has stated.

There is also a more ordinary explanation for the gap than an unfilled scaling wall, and it deserves stating rather than leaving as a mystery. Hypergraph-based tool composition requires annotating a tool's input and output relationships as structure, up front. That is exactly the configuration step Eve's whole philosophy, just files, let a capable model infer the wiring, exists to avoid. That is a mismatch in design philosophy, not necessarily evidence of a problem Eve has not yet hit. Whether Eve's flat approach eventually meets the wall the hypergraph papers are built to solve, many tools, many skills, complex cross-dependencies exceeding what one model context window can plan over, is a real, open, and currently unanswered question. Eve's own generation-three failure, a model that could not generalize to novel questions once given to a wider audience, is a related but not identical wall to the one the hypergraph papers describe. The two are not mutually exclusive, and neither is settled.

#What an operator does with a partial signal

You do not need these two things to be the same story to act on either one. If you are evaluating agent frameworks for your own team, Eve's founding lesson stands on its own regardless of hypergraphs: a team tried four increasingly elaborate architectures before the cheapest-looking one won, and it only won once the underlying model got good enough to be trusted with a general tool surface instead of a prescriptive one. That is a real, falsifiable data point about where complexity is currently worth spending, and it came from someone who lived through the three approaches that lost first.

And if you are watching for what happens when an agent's tools, skills, and memories outgrow a flat structure, the hypergraph research is a real, dated body of work answering exactly that question, worth tracking rather than dismissing as academic. It has not yet found its way into a production framework at Eve's scale. That does not mean it will not. It means nobody has shown that it has, yet, and a wise reader holds that as a question worth watching rather than a verdict to repeat as settled either way.

#Sources

Keep reading