Your curated roundup from genesis-aka.net / IT Professional · 17 articles this week
AI in Software Engineering
How Goldman Sachs Is Using Agentic AI For Software Engineering At Scale (August 13)
Goldman Sachs is running agentic AI software engineers such as Devin alongside its 12,000 human developers to modernize legacy systems, with the agents executing projects independently and cutting bug-fix times sharply. The deployment also raises uncomfortable questions about junior developer redundancy and the resulting skills gap.
At Kilo Code, engineers now spend roughly 1% of their time actually writing code, delegating the rest to AI agents — a shift that has moved the hard problem from productivity to cost control. Replit and Symbotic describe layering human oversight and multi-model routing to keep agent spend from spiraling.
The Agent-Native SDLC and the Verification Bill Nobody Budgeted (August 12)
As autonomous agents take over delivery, verification — not authoring — becomes the bottleneck, at an estimated $900,000 a year for a 50-developer team. The piece argues CI practices must evolve to absorb the elevated risk profile of AI-generated code.
Why Reliability Guardrails Are Needed in Every AI Coding Pipeline (August 12)
Faster AI-driven deployment is translating directly into higher outage risk, and the proposed answer is automated reliability guardrails built into the pipeline. These guardrails test for common failure modes, feed fixes back to the agent, and preserve velocity rather than throttling it.
Developer Tools & Practices
GitHub Brings Stacked Pull Requests Out of the Shadows (August 12)
GitHub has shipped native stacked pull requests, letting developers split a large feature into smaller ordered PRs that review faster and more accurately. The timing is deliberate: as AI tools generate more code, review throughput becomes the binding constraint.
Write Documentation for Yourself, Not Others (August 13)
Reframing docs as a note to your future self makes the chore worth doing — capturing decisions, manual steps, and production fixes while they are still fresh. The payoff is lower cognitive load, easier automation of repeated tasks, and continuity when someone else inherits the system.
Security & Agent Governance
N-central’s Incomplete Patch Left MSP Clients Exposed While Attackers Held Both Keys (August 13)
CISA added three actively exploited vulnerabilities to its Known Exploited Vulnerabilities catalog, including critical flaws in N-able N-central. The bugs let attackers reach both the management server and the downstream endpoints of every organization served by an affected MSP.
AI is finding bugs faster than humans can fix them: How enterprise security teams must adapt (August 12)
AI-driven discovery is generating vulnerability reports far faster than remediation capacity can absorb them. Well-resourced organizations such as Google can keep pace, but most cannot — pushing triage, disclosure policy, and automation to the top of the security agenda.
Rubrik unveils Agent Identity to govern AI agents one tool call at a time (August 12)
Rubrik launched Agent Identity at Black Hat, scoping an AI agent’s permissions on a per-tool-call basis rather than granting blanket access. Paired with Agent Rewind and existing identity system integrations, it targets the blast radius problem created by increasingly autonomous agents.
Cloud & Data Infrastructure
Cloudflare launches Cloudflare OS: an open-source AI agentic workspace for the enterprise (August 13)
Cloudflare has released Cloudflare OS, an open-source AI workspace that lets enterprise employees build custom micro-apps for everyday tasks. The pitch is retained organizational knowledge, full ownership of what teams build, and secure integration with existing systems.
Nvidia open-sources cuFile API, accelerating GPU read/write capability for high-speed storage (August 13)
Nvidia open-sourced its cuFile API and announced a Storage-Next initiative with some 40 technology partners at the Future of Memory and Storage conference. The goal is removing the storage bottleneck that starves GPUs during large-scale AI training and inference.
Latest AI Uses Tabular Foundation Models To Turn Columnar Data Into Vital Insights (August 14)
Tabular Foundation Models are purpose-built for structured data, an area where conventional LLMs reliably stumble. TFMs handle numeric values faithfully and learn patterns across rows and columns, opening up spreadsheet and warehouse data to the same foundation-model treatment text has enjoyed.
AI Frontier & Research
OpenAI’s Astra solves 10 long-open math problems and publishes the proofs (August 11)
OpenAI’s Astra model resolved ten longstanding problems in mathematics and theoretical computer science, publishing machine-checkable proofs and certificates on GitHub — including a non-sofic group construction and a disproof of Connes’s rigidity conjecture. None of the results have yet cleared peer review.
DeepMind Says Its AI Can Predict Hurricanes Earlier Than Everyone Else (August 12)
Google DeepMind’s WeatherNext model called Hurricane Melissa’s track toward Jamaica with 80% confidence, buying forecasters an extra day of warning ahead of the October 2025 landfall. It is a concrete demonstration of learned weather models outperforming conventional numerical forecasting on lead time.
Google’s Top AI Brains Are Leaving to Launch Discovery Loop (August 12)
Jeff Dean spoke to founders at Y Combinator’s Startup School on July 25 shortly before unveiling Discovery Loop, his new venture with fellow Google AI researchers. The startup aims to automate the scientific method itself through machine learning and automated experimentation.
Also This Week
The benefits of medical AI assistance vary based on user expertise (August 12)
An MIT study found AI improves diagnostic accuracy, but unevenly: non-experts over-trust AI explanations and are led astray by faulty ones, while trained clinicians push back more effectively. The lesson is that assistive AI needs to be designed around the user’s expertise level, not deployed uniformly.
Humanoids won’t scale on factory floors until costs drop (August 11)
BMW intends to deploy the Figure 03 humanoid following a successful pilot, but the economics remain the obstacle. Purpose-built specialized robots still beat general-purpose humanoids on efficiency, and meaningful scale awaits lower unit costs and far greater production capacity.
Editor’s Takeaway
This week’s throughline is that the cost of AI in engineering has moved downstream. Writing code is no longer the expensive part — Goldman Sachs runs agents beside 12,000 developers and Kilo Code’s engineers touch code 1% of the time — but the bills are arriving elsewhere: a $900,000-a-year verification burden for a mid-sized team, runaway agent token spend, vulnerability reports piling up faster than anyone can triage them, and reliability incidents from code that shipped faster than it was checked. The responses on offer are all structural rather than clever: reliability guardrails wired into the pipeline, stacked pull requests to make review tractable, Rubrik’s per-tool-call permissions to bound what an agent can actually touch, and better documentation habits so the context survives the handoff. For IT professionals, the practical read is that budget and headcount planning should shift toward verification, review, and governance capacity — the places where agent output accumulates — rather than toward the authoring step that agents have already absorbed.
Explore the full IT Professional archive at genesis-aka.net/information-technology/professional/
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.