Research library / Reviewed 5 September 2026
Search 30 curated sources. Each record identifies what was accessible, what it supports and what it cannot establish. Publisher links open the originals.
Download register (CSV) · JSON · QA companion brief v1.7.0 (13-page PDF)
30 sources
Gartner · 2025-07-01 · Reviewed 2026-09-05
Forecasts 90% of enterprise software engineers using AI code assistants by 2028.
A forecast about tool use, not a measured QE outcome or a savings estimate.
Access: Public text
Gartner · 2026-02-03 · Reviewed 2026-09-05
The public abstract frames platform selection around quality and business value beyond test automation.
Only the abstract was reviewed; detailed criteria, findings and vendor assessments require licensed access.
Access: Public abstract; full report gated
Gartner · 2026-04-02 · Reviewed 2026-09-05
The public summary positions continuous quality as a response to more complex AI-era delivery.
The full guide was not available for review. No proprietary maturity model is reproduced.
Access: Public abstract; full report gated
Gartner · 2026-07-07 · Reviewed 2026-09-05
Forecasts 60% of organizations adopting smaller engineering teams at scale by 2029; emphasizes team redesign and platform enablement.
Not evidence for reducing QE headcount; the release distinguishes restructuring from cost optimization.
Access: Public text
McKinsey · 2025-11-03 · Reviewed 2026-09-05
Surveyed nearly 300 leaders; 100 assessed impact. The top outcome quintile reported stronger productivity, quality and speed.
Self-reported outcomes and selected top performers; no causal estimate and no representative bank savings rate.
Access: Public text
McKinsey · 2026-04-01 · Reviewed 2026-09-05
Describes agent-based delivery and an unnamed global bank implementation as a strategic direction.
Consultancy-reported case; its speed and cost claims lack a public independent counterfactual. Do not transfer them into a business case.
Access: Public text
Capgemini / Sogeti / OpenText · 2025-11-13 · Reviewed 2026-09-05
Reports widespread experimentation, limited enterprise scale and material privacy, integration, reliability and skills barriers.
Vendor-sponsored survey of over 2,000 executives in 22 countries and 10 sectors. Question-specific sample sizes are not in the public release.
Access: Public release; full report registration
DORA / Google Cloud · 2025-09-23 · Reviewed 2026-09-05
AI adoption interacts with the surrounding engineering system; organizational foundations matter for delivery outcomes.
Observational research and reported associations do not isolate the effect of a tool. Partner-sponsored research.
Access: Public text
DORA · 2026-03-10 · Reviewed 2026-09-05
Analysis of 1,110 open-ended Google engineer responses highlights verification work, context limitations and review friction.
One-company qualitative sample from Q3 2025; prompts may have focused respondents on code generation.
Access: Public text
METR · 2025-07-10 · Reviewed 2026-09-05
16 developers and 246 tasks: access to early-2025 AI increased completion time by 19% (95% CI: +2% to +39%).
Experienced developers on familiar repositories; narrow task and tool setting, not a claim about all developers or 2026 agents.
Access: Public text
METR · 2026-02-24 · Reviewed 2026-09-05
Point estimates suggest faster work: returning developers −18% time, new recruits −4%; both confidence intervals include no effect.
Selection effects and concurrent-agent time measurement weaken the estimates. This is not a clean temporal trend against the earlier trial.
Access: Public text
Cui and colleagues / Management Science · 2026-02-27 · Reviewed 2026-09-05
Across 4,867 developers, the pooled estimate was 26.08% more completed tasks, with standard error 10.3 percentage points.
Coding-assistant field experiments, not QE-specific labor savings. Public author PDF is the February 2025 working paper; publication followed in 2026.
Access: Publisher abstract + open author paper
Alshahwan and colleagues / FSE 2024 · 2024-02-14 · Reviewed 2026-09-05
TestGen-LLM improves existing tests using build, reliability and coverage filters before developer review.
Single-company evaluation. Generated tests and test-a-thon recommendations use different denominators; acceptance does not establish defect reduction.
Access: Public text
Meta Engineering · 2025-02-05 · Reviewed 2026-09-05
Uses mutation-guided generation: candidate tests must detect a modeled fault while passing the original code.
Tests target selected fault concerns; effectiveness depends on mutation relevance and human review. Company-authored implementation report.
Access: Public text
Ziftci and colleagues / Google · 2026-04-13 · Reviewed 2026-09-05
Reports 90.14% diagnosis accuracy on 71 manually evaluated failures and deployment on 52,635 distinct failing tests.
The small reviewed accuracy sample is different from the deployment population. Feedback usefulness is not diagnostic accuracy.
Access: Public text
Li and colleagues / UT Austin and Uber · 2025-11-18 · Reviewed 2026-09-05
Uses dynamic call-graph context and analysis to repair flaky tests in a Go monorepo.
Six-month enterprise case. Reported repair and acceptance percentages use different populations; weak fixes can change test semantics.
Access: Public text
NIST · 2024-07-26 · Reviewed 2026-09-05
Provides a cross-sector structure for identifying and managing generative-AI risks throughout the lifecycle.
Voluntary guidance; translate it into system-specific controls and testable requirements.
Access: Public text
NIST · 2023-01-26 · Reviewed 2026-09-05
Connects lifecycle risk management with representative evaluation, documented uncertainty and ongoing monitoring.
RMF 1.0 remains the referenced framework; NIST says it is being revised. It is not a product certification.
Access: Public text
OWASP GenAI Security Project · 2025-12-09 · Reviewed 2026-09-05
Threat guidance extends evaluation to agent goals, tool use, identity, supply chains and memory.
Community risk taxonomy; use it to design scenarios, not to infer measured incident rates for a bank.
Access: Public text
OWASP GenAI Security Project · 2026-08-03 · Reviewed 2026-09-05
The 2026 guide updates LLM application risks and explicitly pairs model-component risks with agentic-system risks.
Resource page dated August 3; release announcement September 1 has a September 2 dateline. Downloaded PDF retains date placeholders. Reviewed as the linked 2026 edition.
Access: Public text
OWASP GenAI Security Project · 2026-09-01 · Reviewed 2026-09-05
Introduces common hooks for runtime policy enforcement and observability across agent frameworks.
Newly donated open standard; evaluate implementation coverage and integration maturity before relying on portability.
Access: Public text
OSFI · 2026-07 · Reviewed 2026-09-05
Discusses scoped agent identities, tool restrictions, testing, traceability, approval and fallback practices for financial institutions.
Technology Risk Bulletin offers sound practices that complement existing guidelines; it is not a new standalone binding AI rule.
Access: Public text
OSFI · 2025-09-11 · Reviewed 2026-09-05
Risk-based model governance and lifecycle expectations provide a planning reference for applicable AI systems.
Effective May 1, 2027. Assess applicability under institutional model policy; do not label every coding assistant a regulated model by default.
Access: Public text
Tricentis · Living documentation · Reviewed 2026-09-05
Documents natural-language support for finding test assets, explaining results and generating test cases.
Capability description for Tosca 2026.1; not an independent assessment of accuracy, enterprise fit or return.
Access: Public text
Applitools · Living documentation · Reviewed 2026-09-05
Describes visual comparison against approved baselines within existing testing frameworks.
Vendor product description. Validate false alerts, missed changes, dynamic content and baseline governance on your applications.
Access: Public text
LangSmith / LangChain · Living documentation · Reviewed 2026-09-05
Documents offline datasets, evaluators and online feedback loops for AI applications.
Tool documentation does not establish evaluator validity. Calibrate judges and human labels to the target task.
Access: Public text
GitHub · Living documentation · Reviewed 2026-09-05
Documents AI review suggestions within pull requests and editors.
Review suggestions supplement reviewer judgment; a product capability is not evidence of prevented defects.
Access: Public text
Promptfoo · Living documentation · Reviewed 2026-09-05
Documents adversarial testing of agent tool use, access boundaries and context.
A red-team suite samples a threat model. Passing it does not prove that an agent is secure.
Access: Public text
Microsoft Foundry · Living documentation · Reviewed 2026-09-05
Documents run-level and row-level evaluation comparison, including quality and operational metrics.
Platform evaluation requires representative datasets, appropriate metrics and deployment-specific thresholds.
Access: Public text
Stack Overflow · 2025 · Reviewed 2026-09-05
Reported trust in AI accuracy remains mixed; distrust exceeds trust in the survey responses.
Self-selected developer survey. Overall sample size and AI-question response counts are different denominators.
Access: Public text
The local archive contains the successfully retrieved public PDFs and a checksum manifest. Gated reports were not downloaded. This site links publisher originals and publishes its own synthesis; it does not rehost third-party reports.
Research coverage, evidence grades and exclusions · Current audience PDF exports