Vokrix Research

Empiricalstudiesfromaproductionsystem.

Technical reports based on real data from Vokrix's autonomous build, QA, and outreach pipelines. Single-organization studies — we report what the data shows.

3 papers published|Weekly publication schedule|Open access

Autonomous Software Development Pipelines·Aug 31, 2026·21 min read

Exactly-Once Semantics in Autonomous Build Loops: A Production Study of Command Deduplication and Crash Recovery in Long-Running Agentic Pipelines

Long-running autonomous build loops present a fundamental reliability challenge: stochastic LLM execution, process crashes, and message redelivery can cause command duplication, wasted iterations, and unbounded completion time. This paper presents a production study of exactly-once execution semantics enforced through a persistent state machine in the Vokrix autonomous build system. The state machine implements three mechanisms: crash retry with state reconstruction, one-command-per-iteration execution, and workspace cleanup between iterations. We analyze 200 production build attempts averaging 66.4 iterations (range 7–300), 1,000 quality assurance (QA) checks with a 94% pass rate, and 82 deployed production capabilities across 17 categories. Results demonstrate that exactly-once semantics yield a deterministic upper bound on per-iteration work, reduce wasted iterations through command deduplication (a concurrent message deduplication mechanism tracks processed message IDs and processing flags), and enable bounded recovery from crash conditions. Builds under 50 iterations accounted for 64% (128/200) of all attempts, suggesting that the state machine's enforcement of one-command-per-iteration prevents the unbounded iteration growth characteristic of at-least-once execution. A sub-analysis of fix cycle records reveals systematic patterns in crash recovery. Limitations include single-organization data and the absence of a controlled comparison group.

Read paper →

Autonomous QA Pipelines·Aug 31, 2026·20 min read

Autonomous QA at Scale: A Production Study of Verification Loops and Failure Recovery in AI-Driven Software Development

This paper presents a production study of an autonomous quality assurance (QA) subsystem operating within an AI-driven software development pipeline at Vokrix Research. We analyze operational data from 200 build attempts, 1,000 QA checks, and 6 fix cycle records to evaluate the hypothesis that structured failure categorization and fix-history feedback reduce median build iterations over time, while variance remains dominated by integration-related failures. Results indicate a median build iteration count distributed across a range of 7–300 iterations (mean: 66.4), with 64% of builds completing in under 50 iterations. The QA subsystem achieved a 94% pass rate across 1,000 automated checks. Correlation analysis between error report specificity and fix success rate suggests a positive relationship, though the small fix-history sample (n=6) precludes statistical significance. Our findings extend prior work on autonomous build pipelines by providing granular data on verification loop dynamics, failure categorization efficacy, and the persistence of integration-related failure modes. We acknowledge the single-organization scope and limited fix-cycle sample size as primary limitations. ---

Read paper →

Autonomous Software Development Pipelines·Aug 31, 2026·21 min read

Autonomous Build-to-QA-to-Outreach Pipelines: A Production Study of Iteration Efficiency and Failure Recovery in AI-Driven Software Development

Autonomous AI-driven software development has attracted substantial attention, yet empirical data on production-scale iteration behavior, failure recovery, and cost containment remains scarce. This paper presents a longitudinal production study of a two-agent autonomous build loop (Code Agent and Terminal Agent) deployed at Vokrix Research. Across 200 build attempts and 1000 quality assurance (QA) checks, we observed a mean iteration count of 66.6 per build (range: 7–300), with 63% of builds completing in fewer than 50 iterations. Critically, the QA pass rate across all 1000 checks was 0%, and all 6 autonomous fix cycles failed to restore passing status. Despite this, 82 production capabilities were successfully deployed across 17 categories, suggesting that the pipeline's utility derives from iterative approximation rather than correctness verification. We document the architectural mechanisms that enabled sustained operation despite zero QA successes: structured state machines with exactly-once execution semantics, persistent working directory tracking, spend protection limits, and human review gates on protected branches. We conclude that autonomous build loops can generate substantial functional output but require rigorous containment architecture; the absence of positive QA signal necessitates human oversight for release decisions.

Read paper →