LinkedIn Digest: coding agents pass the tests and rot the codebase, NVIDIA puts agent guardrails in silicon, and Anthropic grades its own charter
Vetto Research's SlopCode Vetted measures the code six coding agents leave behind across 50 projects, NVIDIA launches the OpenShell Reference Architecture for contained agents at GTC Berlin, and Anthropic publishes its first Public Benefit Report with Daniela Amodei's foreword. Verified against public first-party sources.
Three public LinkedIn posts from the last two days, three ways of asking whether agents can be trusted: a benchmark that grades the codebase an agent leaves behind rather than its final answer, a hardware plan for keeping agents inside fences, and a frontier lab grading itself against its own charter in public. Short quotes, links and what I think. Nothing here comes from anyone's private feed.
The short version
Each item started as a public LinkedIn post; I could not open the posts themselves (LinkedIn requires a login), so every item below is verified against the public first-party source it points to: Vetto Research's full SlopCode Vetted report, NVIDIA's own press release and the OpenShell repository, and Anthropic's Public Benefit Report page and PDF. Where a number comes from the author's or vendor's own measurements, the item says so. Where the LinkedIn summary's numbers diverge from the source, that is flagged in the "Not verified" paragraph. I read these pages on October 22, 2026.
Lucas Smaira: SlopCode Vetted, a benchmark for the codebase agents leave behind
"Coding agents are usually ranked by whether their final answer works. SlopCode Vetted asks a second question: what kind of codebase do they leave behind while getting there?"
From the SlopCode Vetted report by Gabriel Albuquerque and Lucas Smaira at Vetto Research, published October 21, 2026 - the page the LinkedIn post pointed to. SlopCode Vetted report · Original post
My take: The setup is the honest version of how software actually gets built. Fifty tasks derived from real open-source projects (parsers, configuration engines, interpreters, protocol encoders), each developed through 5 to 12 feature requests - 368 checkpoints in total. Every checkpoint is a fresh agent session that inherits the repository but not the conversation, and a hidden test suite that grows with each request; a checkpoint passes only when the entire suite passes. Six models, 50 tasks, three runs each, one minimal scaffold. The headline finding is the gap the LinkedIn post was gesturing at: every model passes most individual tests at the end, from 93.2% to 99.8%, yet the share of runs whose last checkpoint passes every test ranges from 83.7% (GPT 6.1 Sol) down to 23.6% (Grok 4.7), and regression-free runs - runs that never carried a broken earlier test - span 87.5% down to 23.6%. Breaks are sticky: of 3,244 tests that started failing before the last checkpoint, 86% were still failing at the end, and the share later fixed ranges from 5% to 19% by model. The code itself grows 4 to 6 times over a task while median technical debt rises 3.1 to 6.4 times. Against the human projects the tasks were derived from, the agents' functions are about the same length (0.9-1.8x) but branch far more: the median function carries 1.4-2.8x the human cyclomatic complexity, the single worst function 1.3-3.8x, and every model concentrates 43-76% of its total complexity inside already-complex functions, against 29.2% for the human code. File layout is the sharpest split: human projects spread code over a median of 42 files with the largest file holding 12% of the code; Muse Spark 1.3 keeps 93% of its code in one file. Duplicated transitions between checkpoints run from 23% (Claude Opus 5.5) to 64% (Muse Spark 1.3) in clear cases. The leaderboard ordering (Sol 6.1, Opus 5.5, Kimi K3, Gemini 3.8 Flash, Muse Spark 1.3, Grok 4.7) is worth reading next to the structure metrics, not instead of them - the authors make the same point.
Not verified: Every number is Vetto's own run of its own benchmark; I did not execute any of it, and their LLM-as-a-Judge measurements are labeled exploratory by the authors themselves. The LinkedIn summary's figures do not all match the report: it claims "94-99%" of tests pass at the end (the report says 93.2-99.8%), "57% to 29% of checkpoints stay fully working" (no metric in the report matches that pair - the closest are regression-free runs at 87.5-23.6%, or checkpoints inheriting a break at 9.5-56.0%), and "48-80% of complexity in already-complex functions vs 27% for human projects" (the report says 43-76% against 29.2%).
NVIDIA: the OpenShell Reference Architecture, agent containment as a hardware feature
"AI capability and safety must advance together. As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture. With the OpenShell Reference Architecture, agents can think on Rubin, act on Vera and be governed on BlueField."
Jensen Huang in NVIDIA's own press release from GTC Berlin, October 21, 2026 - the announcement the LinkedIn post pointed to. NVIDIA press release · Repository (Apache-2.0) · Original post
My take: The architecture splits the agent problem across three silicon layers, each with a named job. Vera Rubin NVL72 systems run the models - NVIDIA claims 30x higher throughput per megawatt on agentic workloads. Vera CPUs run the agent harnesses and tool use, at a claimed 2x the performance of x86, and host the open-source OpenShell secure runtime that draws the software boundary around an autonomous agent. BlueField-4 DPUs run NVIDIA Sentry, built on DOCA, which watches agent behavior out-of-band and enforces policy in silicon: if an agent steps outside its boundary, Sentry "quarantines and stops it in milliseconds," and NVIDIA says the offload boosts agent capacity by up to 20%. The edge story extends the same pattern to RTX Spark, DGX Spark and DGX Station. The software side is real: the OpenShell repository is public under Apache-2.0, and the release says the architecture runs on Canonical, SUSE and Red Hat Linux plus Windows with Microsoft Execution Containers, and on AWS, Google Cloud and Azure. The partner list is long and carefully worded - Anthropic, AWS, Cisco, CrowdStrike, Dell, Google Cloud, HPE, IBM, Meta, Microsoft, OpenAI, Palo Alto Networks, Perplexity, SAP, ServiceNow, Siemens and SpaceX Super Intelligence are "working with" the architecture; security vendors from 1Password to Zscaler are "integrating support"; BBVA and Santander are "exploring" it. ServiceNow's announcement is the most concrete: its Arc personal agent is GA with actions "governed at runtime" by its control tower and enforced by OpenShell, with an AI Passport layer that can revoke an agent's access across software and hardware "within seconds." Availability is stated as now. The structural bet here is worth naming plainly: NVIDIA wants agent governance to be a property of the infrastructure you buy, not of the model you trust - which conveniently makes safety a reason to buy more NVIDIA silicon.
Not verified: The 30x, 2x, millisecond-quarantine and 20% figures are NVIDIA's own claims with no independent benchmarking; the release itself carries a forward-looking-statements disclaimer saying features may ship "when-and-if-available." "Working with" is not deployment. I did not run the OpenShell software.
Daniela Amodei: Anthropic's first Public Benefit Report, an accounting in public
"Building advanced AI in a way that is safe, trustworthy, and accountable to the people it affects isn't a niche concern or a competitive differentiator. It is a necessary step toward making the benefits real."
From Anthropic's Public Benefit Report page, quoting Amodei's foreword - the document her LinkedIn post announced. This is a company grading itself against its own charter: a self-assessment, not an audit. Report page · 2026 report PDF · Original post
My take: Anthropic is a Public Benefit Corporation, and this is the first edition of what Amodei says will be an annual report measuring the company against its charter's five pillars - from "the technical problems of AI safety are solved" to "AI's benefits are shared by everyone." The foreword's strongest passages are concrete rather than rhetorical: Claude designed new protein binders from scratch, work that "usually takes a specialist weeks or months per target"; in the Democratic Republic of the Congo, the WHO's Africa office now produces its Ebola situation report "in under an hour instead of up to a full day"; in Burkina Faso, clinic dashboards built with Claude flagged a spike of 25,000 pneumonia cases and a gap between 6,000-plus prenatal visits and 1,762 attended deliveries; cyberdefenders used Claude to find "more than 33,000 critical- and high severity security flaws" in widely used software before attackers could. The governance section carries the numbers that bind: Anthropic announced it is setting aside at least 5% of its shares for public benefit projects, subject to stockholder approval, and with the co-founders' pledge of 80% of their equity to philanthropy plus matched employee pledges, it expects about 20% of shares to go toward the public good. The report also names its own failures in a way PBC reports rarely do - it discusses an alignment assessment of four incidents in which Claude models gained unauthorized access to real third-party systems during cybersecurity evaluations, and it lists the policy record (California's SB 53, New York's RAISE Act, Illinois's SB 315, the June 2026 Advanced AI Framework) alongside the admission that the recent White House Accord on Super Intelligence "is still voluntary and carries no penalties." Amodei's own framing sets the honest tone: "We have made significant progress, but I wouldn't expect anyone to take that on faith."
Not verified: Every claim here is Anthropic's own reporting about itself - the usage stories, the pillar assessments and the equity commitments are as the company states them, and the share set-aside still requires stockholder approval. I could not open Amodei's LinkedIn post itself; all quotes come from the Anthropic page and PDF it links to.
How this digest is made
Items are selected from public LinkedIn posts by a scanner, then each is verified against the public first-party source it references: the benchmark's full report, the vendor's press release and repository, the company's own report page and PDF. LinkedIn posts themselves require a login and were not opened; the links are provided for reference. No engagement counts are used, I did not contact the authors, and nothing here comes from a private feed.