📰 Daily Tech Digest - 2026-07-09

🔥 Top Story

Automating cross-repo documentation with GitHub Agentic Workflows

GitHub describes cross-repo documentation automation with GitHub Agentic Workflows, using the Microsoft Aspire team — ten people building dev tools for distributed applications — as the case study. When a feature PR merges in microsoft/aspire, a workflow runs an agent that analyzes the diff and decides whether docs are needed. If so, the agent drafts documentation in a checked-out microsoft/aspire.dev workspace. A safe-outputs handler creates a draft PR with an SME assigned as reviewer, and a comment on the source PR links to the docs. An engineer then reviews and approves. The reported result is 82 feature docs PRs merged with a 44.8-hour median turnaround and no manual reverse-engineering of changes.

💡 Why it matters: If the gap between release and documentation is closed by human reminders, automating just the draft off the merge event already cuts turnaround substantially.

🔗 Read more · GitHub

Kubernetes & Cloud Native

The CNCF Data Storage in Cloud Native AI White Paper

CNCF

The CNCF Technical Advisory Group Infrastructure released a white paper on data storage in cloud native AI in July 2026, titled "Data On Kubernetes – Data Analytics and AI/ML Workloads." It addresses the data bottlenecks that appear when data-heavy, stateful AI/ML workloads move onto cloud native infrastructure, now that deploying them at scale has become a primary enterprise objective. Technologies covered include Apache Parquet, Iceberg, the vector database Milvus and the distributed caching project Fluid. On storage interfaces it covers the Container Storage Interface (CSI), Container Object Storage Interface (COSI) and FUSE CSI drivers. Data pipeline tooling covered includes Change Data Capture (CDC) and Apache Kafka for streaming. It splits the AI lifecycle into three phases with different storage profiles: throughput-focused model training, latency-sensitive inference, and agentic AI with iterative reasoning.

💡 Splitting training, inference and agentic workloads into distinct storage profiles is the useful frame — it justifies re-examining designs that try to serve all three from one storage tier.

Your Laptop Is the New Production Environment

Docker

Docker argues that developer laptops are becoming production environments, because AI agents now take autonomous actions — modifying code, running tests, accessing credentials — rather than only offering suggestions. Its claim is that security models designed for human operators and infrastructure checkpoints are insufficient for governing agents. The products named are Docker AI Governance, Docker Sandboxes, and the Docker MCP Catalog and Toolkit. Runtime governance here means enforcing controls at the execution layer rather than through prompt-based instructions. Concretely, that covers file system access, command execution, network connections, tool usage and credential availability, applied consistently across laptops, CI and production.

💡 The distinction between instructing an agent not to do something and blocking it at the execution layer is the whole point — a policy that lives in the prompt is a request, not a control.

Announcing etcd v3.7.0

Kubernetes

SIG etcd released etcd v3.7.0. New features include RangeStream, which streams large result sets in chunks instead of buffering entire responses, and a keys-only range optimization that reads from the in-memory index only, reducing backend load. Lease operations are faster and more reliable, and bootstrapping from v3store removes the dependency on the legacy v2 store. A protobuf overhaul replaces outdated libraries with fully supported versions. Core dependencies are updated to bbolt v1.5.0 and raft v3.7.0. On performance, the release reports a significant decrease in CPU usage for etcd members compared with v3.6, more efficient large keys-only range requests and improved memory usage predictability.

💡 Replacing full-response buffering with streaming makes this a release worth measuring if large clusters have been suffering etcd memory spikes.

Network boundary for AI agents using NGINX and OpenTelemetry

CNCF

A CNCF blog post shows how to build a network boundary for AI agents using NGINX and OpenTelemetry, prompted by a conversation at a KCD where an attendee said they would not put an agent in their network because "we don't know what that thing" will do. The setup is a single-node Kubernetes cluster running four workloads — NGINX, Ollama, OpenClaw and an OpenTelemetry Collector — in the same namespace on consumer NVIDIA GPU hardware. NGINX acts as a reverse proxy for inbound traffic and a forward proxy for outbound agent requests, with iptables rules forcing all egress through that single path. The native NGINX OpenTelemetry module emits spans for every request, enabling audit logging and integration with observability platforms. Fine-grained, application-aware traffic shaping is implemented through NGINX configuration mapping, demonstrated by blocking everything except nginx.org and duckduckgo.com. Collected spans can be persisted to audit logs or fed into Jaeger, Grafana or SIEM platforms. The post is explicit that this controls network behavior only and does not evaluate agent intent or guarantee safe decision-making.

💡 The usual blocker for putting agents on an internal network is not knowing what they call out to — forcing egress through one path and emitting spans answers exactly that question.

AI & ML

Data for Agents

Hugging Face

NVIDIA released a set of open Nemotron datasets on Hugging Face for agent development. Nemotron-CC-v2 holds 8.79B examples, enhancing Common Crawl for pretraining with synthetic data. Nemotron-CC-Math-v1 holds 190M examples aimed at improving mathematical reasoning through synthetic questions. Nemotron-Personas-USA is a 1M-example synthetic persona dataset grounded in regional demographics. Privasis-USA provides 1.11M privacy-preserving synthetic records spanning medical, financial and legal contexts. The Nemotron Pre-Training Collection spans over 10 trillion tokens across general, code, math and synthetic domains in 15 datasets. Nemotron Post-Training v3 offers millions of post-training samples across multiple domains and data shapes, visualized through an interactive atlas. Nemotron-Personas Global covers ten countries and more than 2.4 billion people with multilingual, region-specific synthetic data aimed at sovereign AI.

💡 Public synthetic personas and privacy-preserving records give a practical option to organizations that could not build test data because real data cannot leave the boundary.

Our approach to government and national security partnerships

OpenAI

OpenAI published its approach to government and national security partnerships along with its National Security Principles. The context is that governments are beginning to use frontier AI systems for increasingly important work, including national security, particularly in cyber defense and biological security where AI can meaningfully advantage defenders. The principles hold that increasingly capable systems must be deployed in ways that reinforce democratic accountability, meaningful human judgment and the rule of law, strengthening democratic institutions rather than concentrating power. They reflect a cross-company effort: OpenAI engaged leading national security expert David Kris to facilitate the process and provide independent judgment, held listening sessions, and involved employees from research and safety through policy and government partnerships. In the past month, as part of its Daybreak cyber defense program, OpenAI established Trusted Access for Cyber partnerships with Australia, Canada, Japan, the Republic of Korea, France, Germany, Poland, the Netherlands and EU institutions such as ENISA, alongside a growing partnership with the UK government around cyber, testing and evaluation. It is taking a similar approach in biosecurity, having announced expanded trusted access to its GPT-Rosalind model for select U.S. government and allied partners supporting public health and biodefense missions.

💡 Trusted Access partnerships already signed across multiple countries including Korea means public-sector AI discussions should assume access tiers and conditions will differ by country.

Separating signal from noise in coding evaluations

OpenAI

OpenAI published an audit separating signal from noise in coding evaluations, finding widespread task issues in SWE-Bench Pro and estimating that roughly 30% of tasks are broken. Accurately measuring model capabilities matters for sound deployment and safety decisions, including those under OpenAI's Preparedness Framework; when evaluations have flaws that affect results, they give a false understanding of capabilities, misrepresenting safety cases and skewing research priorities. OpenAI had previously investigated how SWE-bench Verified, one of the most widely used coding benchmarks, had fundamental design and contamination issues that left it without meaningful signal on software development capability, and encouraged the community to switch to SWE-Bench Pro. SWE-Bench Pro was designed to improve on it by testing models over longer horizons on more realistic tasks to better track agentic coding capability, and on the 731-task public split frontier models improved from a 23.3% pass rate to 80.3% in eight months. The new audit used a datapoint analysis pipeline reviewing model attempts, task metadata and failure traces to flag likely evaluation flaws, with each flagged task then assessed through multiple investigator-agent passes and independently reviewed by five experienced software engineers, escalating disagreements for further investigation.

💡 Two widely used coding benchmarks being found flawed in succession is the takeaway — maintaining a small evaluation set drawn from your own repositories beats selecting models on public scores.

Helping K–12 educators build practical AI skills

OpenAI

OpenAI Academy, in partnership with the Walton Family Foundation, will host a flagship AI Skills Jam for K–12 Educators this summer. The context is that K–12 teachers, administrators, district leaders and school communities help students learn and grow every day while being asked to navigate rapid technological change amid limited time, growing responsibilities and uneven access to practical support. The Jam aims to offer a hands-on, high-trust environment where teachers and administrators can move from curiosity to practical use. Building on the success of the Nonprofit AI Jam and Small Business AI Jam, it will bring together more than 1,600 K–12 teachers, administrators and district leaders for in-person, hands-on workshops across U.S. cities.

💡 The repeated choice of in-person hands-on format is the signal — when designing internal AI enablement, it argues that working sessions convert to actual use better than distributing documentation.

Native-speed vLLM transformers modeling backend

Hugging Face

Hugging Face reports that the vLLM transformers modeling backend now runs Hugging Face models at native vLLM implementation speed through dynamic layer fusion and optimization. On Qwen3 benchmarks it meets or beats native throughput across a 4B single-GPU model, a 32B tensor-parallel model and a 235B MoE configuration. The mechanism uses torch.fx for static analysis of the model graph and manipulates source code via the abstract syntax tree to identify and fuse optimizable patterns. Fused operations map onto vLLM kernels such as MergedColumnParallelLinear and QKVParallelLinear for tensor and expert parallelism. It is enabled with the --model-impl transformers flag and composes with existing tensor, data and expert parallel options. It works with transformers-based models on the Hub; linear attention models are not yet supported and non-compliant custom models are unlikely to work. Optimized models remain torch-compilable and can pass through torch.compile and CUDA Graphs.

💡 It removes the need to wait for or port a dedicated vLLM implementation per model — a direct win for teams that must get new models into serving quickly.

Cloud Updates

C4N, now GA: Delivering cloud’s highest per vCPU network and block storage I/O for x86 workloads

Google Cloud

Google Cloud made C4N, its network and storage optimized VM family, generally available, framed around network and block storage becoming the bottleneck as organizations scale high-throughput databases, network and security appliances, real-time analytics and AI/ML inference. C4N runs on 5th Gen Intel Xeon Scalable (Emerald Rapids) processors. Networking reaches up to 400 Gbps VM-to-VM bandwidth, 95 million packets per second and 200 Gbps internet egress. Storage reaches up to 25 GiB/s throughput and 1M IOPS with Hyperdisk Extreme. Nine machine configurations span 2 to 192 vCPUs with up to 1.5 TB of DDR5 memory. Against C4, Google cites 33% higher network bandwidth per vCPU, 224% faster packet processing and 33% higher storage bandwidth per vCPU, with real-world gains of 1.5x Nginx requests per second and 45% better MySQL QPS for disk-resident data. It is available on-demand, as Spot VMs, reservations and CUDs in us-central1, us-east1, us-east5, us-west1 and europe-west2.

💡 A 224% gain in packet processing is the number that matters for appliances where PPS is the ceiling — firewalls and proxies may need fewer instances.

Google Cloud named Leader in the 2026 Gartner® Magic Quadrant™ for AI Infrastructure

Google Cloud

Google Cloud announced it was named a Leader in the inaugural 2026 Gartner Magic Quadrant for AI Infrastructure, positioned highest for Ability to Execute and furthest for Completeness of Vision. On custom silicon it cites TPU 8t at 9,600 chips per superpod with 3x compute performance, and TPU 8i with 288 GB of high-bandwidth memory and 384 MB of on-chip SRAM. For storage, Managed Lustre delivers 10 TB/s bandwidth — described as 20x faster than competitors — and Rapid Buckets handles 20 million operations per second. The Virgo network connects more than 1 million TPUs or 960,000 GPUs across multiple data centers. GKE Inference Gateway raises throughput up to 40% while cutting serving costs up to 30%. AI Hypercomputer scales to 130,000 nodes at 97% productivity, and GKE Agent Sandbox provisions 300 sandboxes per second. The post says the infrastructure serves nine of ten frontier AI labs, with customers including Citadel Securities and Mercedes Benz.

💡 If inference serving cost is the pain point, the claimed 40% throughput and 30% cost figures for GKE Inference Gateway are a concrete candidate to validate against your own traffic pattern.

New ways to keep Google Cloud certifications current and boost your career

Google Cloud

Google Cloud introduced new ways to keep certifications current. Instead of a traditional proctored exam, candidates can now recertify through Google Skills courses and skill badges. The skill badge path is a fast track using hands-on labs that validate real-world problem-solving, while the course path covers updated courses and labs teaching what has changed. Completing the required activities while the certification is still active automatically extends it by one year. It applies to Cloud Digital Leader, Associate Cloud Engineer, Professional Cloud Architect and Professional Data Engineer. Recertification runs through Google Skills at skills.google.com.

💡 The automatic extension only applies while the certification is still active — noticing after expiry closes this path, which is the condition that actually matters.

Introducing Meerkat: an experiment in global consensus

Cloudflare

Cloudflare Research described Meerkat, an experimental global consensus service for managing control-plane state across more than 330 data centers. It uses a new consensus algorithm called QuePaxa, designed to avoid leader-based bottlenecks so that all replicas can perform writes at all times without timeout-dependent elections. The design goals are linearizable strong consistency and fault tolerance while remaining available as long as a majority of replicas are operational. Against Raft, the claimed advantages are no required leader (removing a single point of failure), concurrent proposals that work constructively rather than conflicting, and a design premised on unreliable, adversarial networks. Intended applications are transactional key-value stores, leasing systems and similar state management needing consistency. Latency depends on replica-to-replica latency, and the system is optimized for infrequently written control-plane data rather than general databases. It is experimental and internal-only, with proofs-of-concept completed at up to 50 global replicas and not deployed to production.

💡 Leaderless consensus is attractive for geographically distributed control planes because it removes election stalls, but the caveat that it targets infrequently written data is what bounds where it applies.

Unleashing open innovation: How Diebold Nixdorf reimagined global banking on Red Hat OpenShift

Red Hat

Red Hat profiles Diebold Nixdorf's global banking platform transformation, presented by Software Solution Architect Joerg Meyer at OpenShift Commons Amsterdam, the "Day zero" event for KubeCon + CloudNativeCon Europe 2026. The company built Vynamic Payments, a microservices-based payments platform using domain-driven design, to replace legacy mainframe systems. It selected Red Hat OpenShift as the runtime for cloud-agnostic deployment across AWS, Google Cloud, Azure and on-premises. Blue-green deployments enable zero-downtime updates, and Kafka-based database replication provides active-active high availability across two data centers. The system scales automatically and keeps processing transactions even during a complete AWS region failure. OpenShift operators integrate third-party dependencies such as Kafka and databases, significantly reducing operational overhead. The legacy systems had relied on expensive proprietary hardware and older Java versions (Java 8), creating security and operational risk.

💡 A claim of surviving a full region failure only holds up with active-active plus a replication story — when evaluating a similar design, examine the consistency boundary of the Kafka-based replication too.

The new currency of enterprise velocity

Red Hat

Red Hat argues about what now constitutes the currency of enterprise velocity, starting from the observation that for more than 20 years enterprise software procurement followed a predictable script: buy a subscription for an open source solution, lock the version and avoid updates unless something broke. It cites that over 90% of modern enterprise codebases rely on open source or third-party dependencies as the reason that premise no longer holds. Red Hat and IBM's response is Lightwell, described as the largest single commitment to open source security since IBM's landmark $1 billion investment in Linux in 1999. Lightwell Network, now generally available, delivers remediated packages with validated, zero-drift fixes to member repositories. Lightwell Clearinghouse Premier provides specialized vertical threat coordination with secure embargo windows before vulnerabilities are weaponized. AI performs high-volume vulnerability triage while human engineers validate fixes to prevent breaking downstream systems. The conclusion is that enterprise velocity now depends on patching capacity rather than static version locks.

💡 The claim that locking versions and leaving them alone is no longer a safety strategy prompts the question of whether patch turnaround is being measured as an organizational capability at all.

Strengthening the open source supply chain with Red Hat partners

Red Hat

Red Hat outlines how it is strengthening the open source supply chain with partners. Earlier the same day, Red Hat and IBM unveiled two commercial Lightwell offerings delivering automated vulnerability remediation at scale, but the post argues that true security requires a movement — a connected network of participants. Technology partners integrating remediation capabilities include AWS, AMD, F5, GitLab, Intel, JFrog, Microsoft, NVIDIA, Palo Alto Networks and ServiceNow. Services partners providing expertise and change management include Accenture, Atos, Cognizant, Deloitte, EY, HCLTech, IBM Consulting, Infosys and Kyndryl. Partners can engage as Lightwell customers, integrate it into their own platforms, or help bring it to market. Lightwell delivers remediated packages plus a core security framework that customers can integrate into products and platforms. Red Hat positions this as the first wave of a scalable, standardized framework for open source security and plans to expand partner engagement into financial services and other critical sectors.

💡 Consuming remediated packages is a transfer of trust — if you evaluate it, the deciding question is whether patch provenance and validation satisfy your own audit requirements.

DevOps & Infrastructure

GitHub availability report: June 2026

GitHub

GitHub published its June 2026 availability report covering six incidents that degraded service. On June 4 at 17:30 UTC, Copilot code review failed for 1h25m (81.6% average, 93.9% peak) after an incompatible dependency version was consumed automatically without validation. On June 8 at 06:30 UTC, unauthenticated users received 504 errors on pull requests, issues and releases for 2h6m (17% average, 34% peak) when abusive automated traffic overloaded the unauthenticated server pool. On June 10 at 15:05 UTC, API authentication failed for 9% of REST and GraphQL requests over 1h20m due to a memcached proxy configuration error causing intermittent lookup failures. On June 16, the Opus 4.8 model was degraded in Copilot for 55 minutes because of an upstream provider issue, with other models unaffected. On June 17, frontier chat models were unavailable across regions for 54 minutes after an invalid configuration change was rejected by the production system. On June 25, the background job service degraded for 23 minutes with peak delays of 7 minutes on PRs, pushes and workflows.

💡 Two of six incidents traced to configuration changes and auto-consumed dependencies — a reminder that config and dependency paths are often less validated than deploy gates.

“Opus-class, but faster”: What Elon Musk says about beating Anthropic

The New Stack

SpaceXAI CEO Elon Musk announced on Wednesday via X that Grok 4.5 would be released publicly on Thursday, citing strong positive feedback from beta program customers and calling it "an Opus-class model, but faster, more token-efficient and lower cost." The public launch came after just under two weeks of internal testing at SpaceX and Tesla. Grok 4.5 runs on xAI's latest 1.5-trillion-parameter V9 foundation model. The model reportedly received specialized supplemental training data from Cursor, the AI coding platform SpaceX is acquiring in a $60 billion deal, which could give it an advantage on software engineering tasks through exposure to high-quality coding workflows and developer interactions. Musk also teased an even larger 2-trillion-parameter V-series model due in August. The "Opus-class" framing references the success of Anthropic's Claude Opus, currently at version 4.8, and the timing is notable: it follows a standoff in which the federal administration imposed strict export controls on Fable 5 over jailbreak-related national security concerns, prompting Anthropic to briefly pull the model before redeploying it with stricter safety classifiers.

💡 Claiming a competitor's tier as a marketing baseline is becoming standard — treat labels like "Opus-class" as unusable for comparison until you measure latency and token efficiency on your own workload.

JetBrains’ next move isn’t a better IDE — it’s a governance layer over Claude Code, Codex, and Gemini CLI

The New Stack

JetBrains announced JetBrains AI for Teams and Organizations on Tuesday. The context is that engineering teams have spent years picking their own AI tools — an IDE here, a terminal-based coding agent there, a browser extension for code review elsewhere — leaving engineering leaders with little visibility into what developers actually use or what it costs. The product sits above whatever AI tools a team already relies on, adding shared context, reusable agentic processes, organization-wide governance and cost controls, rolling out to business customers gradually through July and August. Oleg Koverznev, head of agent systems at JetBrains, wrote that developers reaching for whichever tool suits the task is a freedom worth preserving, and that teams should not have to standardize on a single vendor to benefit from AI. It extends a trajectory that began with a 2023 coding assistant bringing AI into the IDE, followed a year later by Junie, an agent capable of planning and executing its own tasks; in March JetBrains took Junie outside the IDE with a standalone CLI alongside JetBrains Air, an environment for running multiple agents side by side, and Junie left beta in June. Among the four pieces, Automations can trigger cloud agents from repository events or a schedule, running long tasks in managed environments.

💡 Layering governance above tools instead of standardizing vendors gives organizations with mixed coding agents a path to cost and audit visibility without relitigating standardization.

Meta says it caught OpenAI. One thing is missing.

The New Stack

In the same week Mark Zuckerberg told Meta staff the company's AI bets "haven't come to fruition yet," superintelligence chief Alexandr Wang gave a very different message: Watermelon, a model Meta still has in training, had caught up with OpenAI's GPT-5.5 on closely watched benchmarks. Business Insider reported Wang's claim from two people in the room; Reuters heard a recording of Zuckerberg's admission. The article's point is that Wang named a rival, a model and a verdict, but skipped the one thing that would let anyone verify it — the benchmark itself. The claim is single-sourced and internal; Meta declined to comment and OpenAI did not respond, and the two people who described the town hall could not say which benchmarks Wang meant. All that exists is the codename Watermelon, successor to Avocado, the internal name for the Muse Spark model Meta shipped in April. There is no model card, no evaluation harness, no ship date and no table. Muse Spark scored well on standard tests but still trailed OpenAI and Anthropic overall. When a vendor publicly declares parity, a benchmark table typically accompanies the claim — OpenAI ran one with GPT-5.5 and Anthropic runs one with every Opus release.

💡 The point that a parity claim without a benchmark table is closer to an internal morale number is a practical filter — make disclosed reproduction conditions the first gate in any model adoption decision.

How GitHub Copilot enables zero DNS configuration for GitHub Pages

GitHub

GitHub walks through using Copilot CLI to configure DNS for GitHub Pages without editing a single record by hand. Going from an empty repository to a live custom domain with HTTPS took about 14 minutes (11:21 a.m. to 11:35 a.m. ET). The pieces are GitHub Copilot CLI, the Namecheap skill from the awesome-copilot repository, and the Namecheap API. Step one creates a public repository and enables Pages with a generated landing page. Step two registers an inexpensive domain — the example ghpagesblog.click cost $2.00. Step three enables Namecheap API access, installs the skill via gh skill install, supplies credentials and configures the custom domain DNS records. Step four verifies domain resolution and an HTTP 200 response from the deployed site. Copilot CLI requests approval before applying changes.

💡 The approval prompt before the agent mutates external state like DNS is the safety property that matters — the question when widening automation is whether that gate survives.

Multi-Agent Collaboration on a Shared Canvas

Honeycomb

Honeycomb published part 2 of a series on taking agentic applications from prototype to production with AWS AgentCore; part 1 covered foundational design questions such as cardinality, session lifecycle and state management. The architecture maps one Runtime session to one investigation rather than to one user, hosting multiple agent sessions with separate LLM contexts. AWS AgentCore provisions isolated microVMs for each new session. Honeycomb's Canvas is the shared collaborative investigation environment where humans and agents work as peers on a common surface. A coordination plane tracks hypothesis claims, agent activities, findings, peer responses and event queues, enabling real-time inter-agent awareness. Investigations come in two forms: directed, with a single hierarchical orchestrator, and cooperative, where peer agents make independent decisions informed by shared context. The organizing idea is "independence with awareness" — each agent operates autonomously while accessing peers' findings without centralized control.

💡 Scoping a session to an investigation rather than a user is the load-bearing decision — a concrete precedent for where to draw the shared-state boundary when adding multiple agents.

GitLab Patch Release: 19.1.2, 19.0.4, 18.11.7

GitLab

GitLab released patch versions 19.1.2, 19.0.4 and 18.11.7 for Community and Enterprise editions on July 8, 2026, fixing eight security vulnerabilities: CVE-2026-6896 (XSS, CVSS 8.7), CVE-2026-13320 (HTML injection, 7.3), CVE-2026-11827 (credential exposure, 4.9), CVE-2026-8472 (access control, 4.3), CVE-2026-7492 (authorization, 4.3), CVE-2025-12506 (reference ambiguity, 3.5), CVE-2026-13151 (group settings, 2.7) and CVE-2026-6352 (compliance violation, 2.7). GitLab advises that all self-managed installations be upgraded to one of these versions immediately. The release also carries multiple bug fixes, including OAuth application fixes and a Go version bump to 1.25.11. Database migrations are included, so single-node instances will experience downtime during the upgrade, while multi-node instances can apply the patch without downtime using zero-downtime procedures. Post-deploy migrations are available for 19.1.2 and 19.0.4.

💡 With a CVSS 8.7 XSS in the set, self-managed GitLab operators should pull the upgrade forward rather than waiting for the next scheduled maintenance window.

How we used AI agents to migrate GitLab rate limiting

GitLab

A small GitLab team ran an experiment on whether AI agents could migrate part of its legacy rate-limiting system without lowering the safety bar, and concluded yes. The work unified two systems — the application-level Gitlab::ApplicationRateLimiter with 121 keys and a Rack-level system — into a single labkit-ruby implementation. Agents drafted specifications, implemented bounded changes, wrote tests and pre-reviewed MRs inside a strict loop requiring spec review, adversarial review cycles capped at two rounds, implementation, code review and human approval before merge. Deliverables were 14 numbered specifications and over 30 merge requests to labkit-ruby, migrating 95 call sites (83 in the monolith, 12 in EE). Six cohorts rolled out from May 4 to mid-June 2026, reaching 100% coverage of all 121 keys. Safety measures included a gradual 1%→10%→50%→100% rollout, shadow-mode comparison before enforcement, GitLab Duo Code Review on all MRs, and an audit confirming the legacy path was near zero. Shadow mode caught a structural collision where three String values were squeezed into two primitive slots, dropping the identifier on unauthenticated paths, fixed within two days. Initial cohorts missed 17 EE-only rate limits, requiring a sixth cohort to capture the orphaned keys.

💡 Shadow-mode comparison is where the real defect surfaced, showing that in agent-driven migration the basis for trust is comparison against production traffic rather than review.

This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.