AI & Computingpreprint2026-08-11

Weft: Evidence-Gated Version Control for Autonomous Agent Swarms

Open access0 citations

Abstract

Version control systems assume that a human reads each change before it is integrated. Commit messages are prose for a reader; pull requests exist to partition work into human-reviewable units; branch protection encodes “someone approved this.” Autonomous coding agents violate that assumption by roughly two orders of magnitude in throughput, leaving practitioners to choose between throttling agents to human reading speed and abandoning pre-integration verification altogether. This paper presents Weft, a version-control and coordination protocol in which integration is gated by machine-checkable evidence bound to exact content rather than by human attention. Weft contributes: a change object carrying an explicit read-set, enabling detection of semantically stale work whose textual footprint does not conflict — a failure class three-way textual merge cannot observe; evidence bound to a Merkle materialization manifest, so an attestation refers to specific bytes rather than a mutable reference; capability-based agent identity with authorization frozen at certification time, so credential expiry and revocation cannot retroactively invalidate history; and a batching-and-bisection gate that amortizes verification across commutable work while isolating faults. A reference implementation in approximately 7,000 lines of Rust is evaluated on four workloads spanning code, structured data, and prose. Across all four, output assembled without a gate fails its own validator while gated output passes; a 50-agent, 100-task workload lands 82 changes in 15 certified landings, refusing 8 of 8 stale-read changes, 8 of 8 seeded defects, and 3 of 3 revoked credentials. The specification underwent adversarial review by two frontier language models, an executable prototype, continuous integration, and public critique, yielding 92 dispositioned findings — including one fatal content-model defect that fuzzing and model review both missed and only a live demonstration exposed.

// Source

View paper (DOI)Open access versionOpenAlexZenodo (CERN European Organization for Nuclear Research)Published 2026-08-11

Authors: Pranab Sarkar