Do LLM Agents Coordinate on Shared Records? Two Pre-registered Studies of Optional Protection on a Record API

When several LLM agents act on the same business records, a tool API can offer concurrency protection and leave its use to the agents. We built a testbed of versioned records in which a seeded scheduler makes the agents' reads overlap. A ledger scorer, with no LLM judge, checks whether some serial order of the committed agents' intents reproduces the final store. Our internal reviews found that the harness of our first pre-registered study (6 models) decided several outcomes, so we report that study as registered and replicated the question. Study 2, registered before any paid call, used a corrected harness and gave every task a matched control in which each agent has a private copy of the records. For its three confirmatory models, contention added integrity failures under every config that offered protection (S2-H1). The pooled differences were 0.38 with conditional writes, 0.61 with locks and 0.30 with both; no test compares the configs. Under conditional writes, all contended failures were on the three tasks whose constraint spans records (exploratory). No single-record episode failed there, though 22 Haiku 4.5 agents reported sales they never wrote, which this measure does not score. Post hoc, nearly all lock failures came from three mechanisms: locking after reading, locking only the written record when the constraint spans records, and not locking. Most half-done transfers (32 of 35) came from both legs sent in one reply (post hoc). In these instances both agents move the same amount, and every such agent reported success, 21 of them after list_effects had shown no debit. Offered both primitives, the confirmatory models attempted fewer locks (S2-H2a) and mostly used the version check (descriptive). The design made the main confirmatory result likely, so the descriptive and post hoc mechanism results are the main contribution. Files in this record gopnal-swamy-2026-llm-agents-shared-records-v1.pdf: the paper (28 pages). gopnal-swamy-2026-llm-agents-shared-records-v1-supplement.pdf: the supplement (71 pages). coordination-shared-records-v2.zip: the code and data (artifact v2): the code of both harnesses, every transcript and ledger, and the frozen analyses with their outputs. Its content is identical to commit 38eb061 of github.com/gssanjana4/concurrency-bench. Start with its README.md. paper-source-v1.zip: the LaTeX source of the paper and the supplement, the paper-pipeline scripts and fact sheets that are not in the code and data zip, and the report of the verification of the Study 2 results. Pre-registrations. Study 1: osf.io/a5vec (registered 31 October 2026). Study 2: osf.io/zj3an (registered 9 November 2026). Both registrations are embargoed on the OSF until 1 November 2027; the registration texts and their file hashes are in the code and data zip. Licences. Paper, supplement, paper source and data: CC BY 4.0. Code: MIT (LICENSE in the code and data zip). Model outputs are subject to the providers' terms. Preprint, not peer reviewed.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-14
DOI
https://doi.org/10.5281/zenodo.24226164
Primary Topic
Distributed systems and fault tolerance
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Do LLM Agents Coordinate on Shared Records? Two Pre-registered Studies of Optional Protection on a Record API

Sanjana Gopnal Swamy
Zenodo (CERN European Organization for Nuclear Research)
Distributed systems and fault tolerance
preprint

Do LLM Agents Coordinate on Shared Records? Two Pre-registered Studies of Optional Protection on a Record API

Sanjana Gopnal Swamy
preprint en

Abstract

When several LLM agents act on the same business records, a tool API can offer concurrency protection and leave its use to the agents. We built a testbed of versioned records in which a seeded scheduler makes the agents' reads overlap. A ledger scorer, with no LLM judge, checks whether some serial order of the committed agents' intents reproduces the final store. Our internal reviews found that the harness of our first pre-registered study (6 models) decided several outcomes, so we report that study as registered and replicated the question. Study 2, registered before any paid call, used a corrected harness and gave every task a matched control in which each agent has a private copy of the records. For its three confirmatory models, contention added integrity failures under every config that offered protection (S2-H1). The pooled differences were 0.38 with conditional writes, 0.61 with locks and 0.30 with both; no test compares the configs. Under conditional writes, all contended failures were on the three tasks whose constraint spans records (exploratory). No single-record episode failed there, though 22 Haiku 4.5 agents reported sales they never wrote, which this measure does not score. Post hoc, nearly all lock failures came from three mechanisms: locking after reading, locking only the written record when the constraint spans records, and not locking. Most half-done transfers (32 of 35) came from both legs sent in one reply (post hoc). In these instances both agents move the same amount, and every such agent reported success, 21 of them after list_effects had shown no debit. Offered both primitives, the confirmatory models attempted fewer locks (S2-H2a) and mostly used the version check (descriptive). The design made the main confirmatory result likely, so the descriptive and post hoc mechanism results are the main contribution. Files in this record gopnal-swamy-2026-llm-agents-shared-records-v1.pdf: the paper (28 pages). gopnal-swamy-2026-llm-agents-shared-records-v1-supplement.pdf: the supplement (71 pages). coordination-shared-records-v2.zip: the code and data (artifact v2): the code of both harnesses, every transcript and ledger, and the frozen analyses with their outputs. Its content is identical to commit 38eb061 of github.com/gssanjana4/concurrency-bench. Start with its README.md. paper-source-v1.zip: the LaTeX source of the paper and the supplement, the paper-pipeline scripts and fact sheets that are not in the code and data zip, and the report of the verification of the Study 2 results. Pre-registrations. Study 1: osf.io/a5vec (registered 31 October 2026). Study 2: osf.io/zj3an (registered 9 November 2026). Both registrations are embargoed on the OSF until 1 November 2027; the registration texts and their file hashes are in the code and data zip. Licences. Paper, supplement, paper source and data: CC BY 4.0. Code: MIT (LICENSE in the code and data zip). Model outputs are subject to the providers' terms. Preprint, not peer reviewed.

Zenodo (CERN European Organization for Nuclear Research)
Distributed systems and fault tolerance
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon