Confining Nondeterminism: AI-Driven Research Systems as DBMSs for Reliable, Non-Wasteful, Transparent, and Collaborative Research

Jul 16, 2026

Upload by Levi

LLM agents that conduct research can transform a research question into code, experiments, and figures, yet remain difficult to trust and costly to run. Re-asking the same question yields different outputs, reported numbers may not match any execution, and failed steps cannot be reliably undone. The work argues that the core issue is the stochastic nature of each agent loop call. It proposes a DBMS-inspired design: a deterministic, versioned dataflow engine executes results, while the LLM acts only as a stochastic compiler editing the plan, with reliability, efficiency, transparency, and collaboration ensured by database guarantees.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Introduction

Trust failures in current research agents

Cost and inefficiency in the agent loop

The proposed DBMS-style boundary

Confining Nondeterminism: AI-Driven Research Systems as DBMSs for Reliable, Non-Wasteful, Transparent, and Collaborative Research [Vision] Kyoungmin Kim EPFL Lausanne, Switzerland [kyoung-min.kim@epfl.ch](kyoung-min.kim@epfl.ch) Anastasia Ailamaki EPFL Lausanne, Switzerland [anastasia.ailamaki@epfl.ch](anastasia.ailamaki@epfl.ch) arXiv :2607 . 10508v1 [ cs .DB] 11 Jul 2026 Abstract LLM agents that conduct research (proposing ideas, writing and running code, analyzing results) can already carry a study from research question to figures, yet cannot be fully trusted. The same question asked twice in a row returns different answers; the agent announces a number that no execution produced, and having tools does not prevent this, because nothing binds what the agent reports to what its tools returned; a small upstream change leaves downstream results silently stale, with no way to list which ones; and the agent re-runs preprocessing and re-writes code it has already produced, because nothing tells it that an identical step has already run, at the researcher’s expense. We argue these failures share one root: every step of today’s agent loop is a stochastic LLM call whose internal state nobody, including the agent, can check. Rather than trying to see inside the LLM, we take a lesson from databases, which earn trust without being watched, because deterministic operators over well-defined state make their guarantees hold by construction. We propose organizing a research project the same way. The project itself is a deterministic, versioned dataflow engine (in effect, a query plan over materialized views), and the LLM, together with the user, is a stochastic compiler that may only edit that plan. The executor never calls the LLM; LLM output enters only as versioned code and data that the executor then runs; and any result the agent asserts enters the project’s stored record only with an execution behind it. Five design rules at this boundary turn familiar database machinery into guarantees for research. Versioning, idempotence, rollback, provenance, evidence-backed results, and isolation make the system reliable; incremental maintenance, memoization, cost-based scheduling, and approximation make it non-wasteful; an explicit, queryable plan makes it transparent; and because the whole project is now a small database instance, sharing it makes research collaborative, down to peer review conducted as read-only queries over a submitted instance. We map each recurring pain of research agents to the database technique that removesit, show that prior agent–data-systems work covers only the performance side of this map. This report presents the diagnosis, the requirements, and the design; the guarantee walkthrough, a prototype with early results, and the research agenda appear in the full version, in preparation. The LLM, we argue, should be the query compiler, never the executor. Keywords AI-driven research systems, agentic data systems, determinism, dataflow, materialized views, incremental view maintenance, MVCC, provenance, transactions, reproducibility, collaboration, peer review 1 Introduction A new kind of system has begun to do research automatically. Given a goal, say “find out whether data augmentation improves this classifier,” an LLM-based agent proposes an approach, writes the code, runs the experiments, reads the results, and iterates [7, 11, 14] . We call such a system an AI-Driven Research System (ADRS).1 The database community has started to take these agents seriously as a workload; recent work redesigns data systems so that agents can probe them efficiently [6, 10] . This paper asks the opposite question: if we want to trust what the agent produces, and afford to run it, what should the agent look like? Our answer is that it should look like a database system. Two ways they disappoint. The first is that you cannot trust them. Ask the same question twice in a row, with nothing changed in between, and t

FAQs

Why can’t LLM-based research agents be fully trusted today?

Repeated questions can produce different answers, and reported metrics may not correspond to any actual execution. Tool use doesn’t eliminate the gap because there is no binding between what the agent reports and what its tools returned.

What is the paper’s main root cause of these failures?

Each step of today’s agent loop is a stochastic LLM call whose internal state cannot be checked by anyone, including the agent itself. This makes it hard to ensure consistency, correctness, and traceability across the workflow.

How does the proposed DBMS-inspired design improve reliability and collaboration?

It separates responsibilities: a deterministic, versioned dataflow engine executes results, while the LLM and user only edit a stored plan. Results enter the project record only after execution, enabling versioning, provenance, evidence-backed outcomes, and read-only query-based peer review for collaboration.

Jul 16, 2026

Upload by Levi