Morning Brief: Thursday, November 19

Sixty-six feeds reporting. Two weeks. 5,799 items reduced to what follows. (what we track, how we crawl, subscribe)

Google launched Gemini 4 Neon, and the load-bearing number in the coverage is not a weights number: The New Stack reports the coding score tripled without the model changing. The research crop arrived the same morning treating the harness as the thing under engineering.

Two days earlier the same outlet benchmarked a 5x cost spread across harnesses running the same model. Today arXiv has Agentic Harness Engineering Beta evolving harnesses against requirements, BaseRSI folding harness behavior back into base model weights, Where the Tokens Go accounting for heat and error inside agentic coding loops, and Reducing Inference Cost of Coding Agents with Timeout Cascades treating the scaffold as the cost surface. Yesterday The New Stack argued verification, not generation, is where coding ROI lives.

The same week put four named advisories into the runtime layer: Java4Shell, reported as remote code execution in OpenJDK via Base64; nginx2root, carried by LWN as remote code execution in nginx; Ruby2Shell; and an out-of-cycle PHP critical release. Bank CEOs and the NSA told Congress that AI-driven patches are behind recent outages. On the research side, Idempotency Is Not Recovery fault-injects agent tool actions, Dependency Fabrication in Repository Deployment Agents names a supply-chain failure specific to agents, and Intent Contracts proposes taking tools away from them.

Top (5-7 min)

Google's Gemini 4 Neon tripled its coding score without changing the model
The New Stack, 2026-11-19. The gain is attributed to the Antigravity harness, not the weights. If that holds, benchmark comparisons across labs are measuring scaffolds as much as models.
Google launches Gemini 4 Neon, touting lower costs for complex AI coding tasks
InfoWorld, 2026-11-19. The launch framing is cost-per-complex-task rather than peak capability. Third frontier coding model in three days, after GPT-6.2 and MAI Code 2.
Agentic Harness Engineering Beta: Requirement-Centered Evolution of Agent Harnesses
arXiv cs.AI, 2026-11-19. Harness construction treated as its own engineering discipline with requirements and evolution, rather than as prompt plumbing around a model.
BaseRSI: Recursive Self-Improvement of a Base Vision-Language Model via Training-Free Harness Internalization
arXiv cs.AI, 2026-11-19. The reverse direction: distill what the scaffold does back into the base model. Makes the harness/weights boundary a design choice rather than a fixed line.
Where the Tokens Go: Heat, Entropy and Error in Agentic Coding
arXiv cs.AI, 2026-11-19. Token-level accounting of where agentic coding loops spend and where they go wrong. The measurement companion to the harness-cost benchmarks.
Idempotency Is Not Recovery: A Fault-Injection Study of Language-Model Agent Tool Actions
arXiv cs.AI, 2026-11-19. Idempotent tools are the standard answer to agent retries. This injects faults and reports that idempotency does not give you recovery.
Intent Contracts: Taking Tools Away from Agents
arXiv cs.AI, 2026-11-19. Restriction as a control primitive rather than a limitation. Complements RefinementGuard on hidden agent-tool execution in the same batch.
Ruby2Shell (GHSA-34m4-m9fh-gqp5)
Lobsters, 2026-11-19. Third named RCE in the week's language-runtime run, after Java4Shell and nginx2root. PHP shipped a critical release out of cycle the same day.

Themes

The harness is the artifact

Three frontier coding models reached the feeds in three days: GPT-6.2 Astra and Sol on Tuesday, Microsoft's MAI Code 2 the same day, Gemini 4 Neon this morning. The differentiating claims are not all about weights. Gemini 4 Neon's coding score is reported as tripling on harness changes; MAI Code 2 is pitched at near-Opus performance for 5% of the cost; The New Stack's Tuesday benchmark found a 5x cost spread attributable to harness choice alone.

The research side is now populated enough to read as a field. Agentic Harness Engineering Beta puts requirements and evolution around harness construction. GRAFT searches agent designs by execution-guided genetic recombination. BaseRSI internalizes harness behavior into base weights, which inverts the question. SCAFFOLD asks how to transport the warrant for a skill, not just the skill. Semantic Context Handoff and ARISE attack memory across long horizons; Resolution as Attention for Efficient Agent Memory Perception attacks its cost. Reducing Inference Cost of Coding Agents with Timeout Cascades and Where the Tokens Go treat the loop as an economic object.

Your AI coding agent could cost 5x more depending on the harness, benchmark finds
The New Stack, 2026-11-17. The number that makes the rest of this coherent. Same model, different scaffold, five times the bill.
Why verification, not generation, is your secret to AI coding ROI
The New Stack, 2026-11-18. Argues the bottleneck moved downstream of generation. Pairs with the Claude Code team's own write-up on verification.
How the Claude Code team thinks about verification
Pinboard, 2026-11-17. First-party account of where a shipping agent puts its checks.
GRAFT: Execution-Guided Genetic Recombination for Agent Design
arXiv cs.AI, 2026-11-19. Agent design as a search problem with execution as the fitness signal.
Reducing Inference Cost of Coding Agents with Timeout Cascades
arXiv cs.AI, 2026-11-19. Cost control by giving up earlier and escalating, rather than by shrinking the model.
SCAFFOLD: Toward Warrant-Preserving Evidence Transport for Reusable LLM-Agent Skills
arXiv cs.AI, 2026-11-19. Reusing a skill means moving the reason it worked, not only the procedure.
Semantic Context Handoff with On-Demand Evidence for Long-Horizon Agents
arXiv cs.AI, 2026-11-19. Context handoff framed as evidence retrieval rather than summarization.
Building the Agent Host Protocol for sessions that continue across devices
GitHub Engineering, 2026-11-17. The harness-as-product version: session continuity as protocol.

The scaffold is also the attack surface

Four named runtime advisories landed inside a week. Java4Shell (CVE-2026-114938) is reported as reaching OpenJDK through Base64; nginx2root is carried by LWN as remote code execution in nginx; Ruby2Shell is a Ruby advisory distinct from Tuesday's Regexp and String case-conversion vulnerabilities; PHP shipped a critical release out of cycle. Red Hat Quay separately allowed reading and overwriting every private image in a registry.

Running alongside: attackers hijacked Microsoft's agent-framework package to steal Azure and CI credentials, and bank CEOs and the NSA told Congress that AI-driven patches are implicated in recent outages. The research in the same batch is specifically about agents as the failure vector: Dependency Fabrication in Repository Deployment Agents, AutoPI and PIPETTE on prompt injection, Probing agent overeagerness in realistic settings with continuation audits.

Java4Shell: Write once, RCE everywhere
Pinboard, 2026-11-18. The writeup behind the week's largest patch cycle.
How Cloudflare WAF helps protect against Java4Shell (CVE-2026-114938)
Cloudflare, 2026-11-18. Mitigation while the ecosystem patches. Includes the detection signature.
Java4Shell is a reason to use Java
Lobsters, 2026-11-18. The contrarian read, on the grounds that the fix propagated through a coordinated ecosystem.
nginx2root
Lobsters, 2026-11-18. Remote code execution in nginx, with LWN carrying the advisory the same day.
PHP 8.6 Release Announcement
Lobsters, 2026-11-19. Shipped against the backdrop of Wednesday's out-of-cycle critical security notice.
Security advisories: multiple vulnerabilities in Regexp and String case conversion
Lobsters, 2026-11-18. Separate from Ruby2Shell, same week.
Attackers hijack Microsoft's agent-framework package to steal Azure and CI credentials
InfoWorld, 2026-11-18. Agent tooling as the distribution channel for credential theft.
Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages
Slashdot, 2026-11-18. Testimony, not analysis. Notable because it attributes outages to automated remediation rather than to the vulnerabilities.
Dependency Fabrication in Repository Deployment Agents
arXiv cs.AI, 2026-11-19. Hallucinated dependencies as an exploitable surface in deployment automation.
Critical Red Hat Quay Flaw Lets Attackers Read and Overwrite All Private Container Images in a Registry
Lobsters, 2026-11-17. Registry-wide read and write from one flaw.

Agents as observability consumers

Two independent items frame agents as the primary readers of telemetry rather than its subject. Groundcover's CEO argues agents are observability's power users; a Hackaday build gives a coding agent audio input to localize a mystery beep. StatePortal synthesizes checkpoints to test deep states in interactive software, which is the same move applied to test infrastructure.

Why Groundcover's CEO says agents are the power users of observability
The New Stack, 2026-11-19. Inverts the usual framing. The consumer of the dashboard is the thing being debugged.
Tracking Down a Mystery Beep By Giving a Coding Agent a Few Ears
Hackaday, 2026-11-19. Instrument first, then let the agent read the instrument. A small worked example of the pattern above.
StatePortal: Agentic Checkpoint Synthesis for Testing Deep States in Interactive Software
arXiv cs.AI, 2026-11-19. Getting a test to a deep state is usually the expensive part.
Find the log line that matters faster with full-text search in Grafana Cloud Logs
Grafana Labs, 2026-11-17. Full-text search over logs, which is also what an agent needs.

Scan

Software engineering and coding agents

Metrics-in-the-Loop: Complexity Feedback Reshapes Agent-Written Code
arXiv cs.AI, 2026-11-19. Feeding complexity metrics back changes what the agent writes, which makes the metric a design input rather than a report.
When Single Agent Is Enough: Rethinking Multi-Agent Systems for Automated Code Review
arXiv cs.AI, 2026-11-19.
Repository-Grounded Rubrics for Execution-Free Evaluation of Coding Agents
arXiv cs.AI, 2026-11-19.
When Code Works but the License Doesn't: Evaluating License Constraint Handling in Coding Agents
arXiv cs.AI, 2026-11-19.
The Obsolescence of High-Level Code: Agentic Generation from Semantic Artifacts
arXiv cs.AI, 2026-11-19.
3D GameDev Arena: Benchmarking Coding Agents That Plan, Build, and Playtest Long-Horizon Browser Games
arXiv cs.AI, 2026-11-19.
Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization
arXiv cs.AI, 2026-11-19.
Maintainers don't need more AI bug reports. They need fixes.
GitHub Blog, 2026-11-17. The supply side of agent output, from the people absorbing it.
Microsoft's new coding model promises near-Opus performance at 5% of the cost
The New Stack, 2026-11-17.
Microsoft launches MAI Code 2 models to challenge OpenAI and Anthropic in AI coding
InfoWorld, 2026-11-18.
External sandboxes for coding agents
Simon Willison, 2026-11-17.
Inside the Codex sandbox
OpenAI, 2026-11-18.
Accelerating development for massive Git repositories
Lobsters, 2026-11-18.
Stacked diffs with Maiao
Lobsters, 2026-11-18.
Snapchat now lets developers build AR Lenses with AI coding assistants like Claude Code and Cursor
TechCrunch, 2026-11-18.
claude-code v2.1.328
GitHub releases, 2026-11-19.

Evaluation, benchmarks, and interpretability

Can you make models believe false facts in the prompt?
arXiv cs.AI, 2026-11-19.
Probing agent overeagerness in realistic settings with continuation audits
arXiv cs.AI, 2026-11-19.
Choosing What to Observe: Budget-Paired Counterfactual Audits for Credible Evaluation of Visual Agents
arXiv cs.AI, 2026-11-19.
Training Language Models for Introspective Oversight
arXiv cs.AI, 2026-11-19.
How Complete Must an Environment Be to Train a Capable GUI Agent?
arXiv cs.AI, 2026-11-19.
WorkWorlds: Evaluating Agents in a Company Built Before the Task
arXiv cs.AI, 2026-11-19.
MANGO Arena: Same Seeds, Different Chains Evaluating LLM Agents in a Disrupted Supply Chain
arXiv cs.AI, 2026-11-19.
Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems
arXiv cs.AI, 2026-11-19.
Openhagen Worlds: Calibrated Environments for Measuring and Training Scientific Discovery Agents
arXiv cs.AI, 2026-11-19.
A Capability Framework for Evaluating Agentic AI Platforms in Regulated Enterprises
arXiv cs.AI, 2026-11-19.
Evaluating Governed LLM-Driven ML Experimentation: Lifecycle, Provenance, and Failure Metrics
arXiv cs.AI, 2026-11-19.
Learning Trajectories Leave Decodable Traces in Neural Network Weight Space and Internal Representations
arXiv cs.AI, 2026-11-19.
Modeling LLM Reasoning Dynamics with Structural Representations
arXiv cs.AI, 2026-11-19.
BOS Summary: The First Token Matters
arXiv cs.AI, 2026-11-19.
AINews: Gemini 4 Neon, GDM's Opus/Sol Contender
Latent Space, 2026-11-19.
AINews: GPT-6.2 Astra is better and 20-600% cheaper, but has mixed reviews
Latent Space, 2026-11-18.

Infrastructure, outages, and the buildout

Traffic mix matters
Marc Brooker, 2026-11-18. Why benchmark numbers taken under a uniform workload mislead.
Outage Chaos: When Tech is too Big to Fail
EFF Deeplinks, 2026-11-18.
Microsoft's outage should terrify every enterprise
InfoWorld, 2026-11-17.
Microsoft Blames Monster Email Outage On Software Defect Introduced In July Fix
Slashdot, 2026-11-16. A fix that caused the outage four months later, which is the same shape as the AI-patch testimony.
Incidents in the news
Lobsters, 2026-11-18.
Satya Nadella says lots of critical infrastructure is 'unpatchable'
TechCrunch, 2026-11-18.
Will agents make Kubernetes obsolete?
InfoWorld, 2026-11-17.
Kubernetes Failure Stories
Pinboard, 2026-11-17.
Microsoft's ARM-based Cobalt 200 chips go GA. Why it matters for AI agents.
The New Stack, 2026-11-17.
CoreWeave connects custom training loops to model serving
The New Stack, 2026-11-17.
Netmap caching: keeping Tailscale up on iffy connections
Tailscale, 2026-11-17.
Quad9's withdrawal from France: A canary in the coal mine
Cloudflare, 2026-11-17.
Codeberg's domain suddenly disappeared from their account, site is unreachable
Lobsters, 2026-11-18.

Clojure and the Lisps

Clojure 1.12.7
Planet Clojure, 2026-11-18.
clojure-1.12.7 release
GitHub releases, 2026-11-18.
Sexp Matrix
Lobsters, 2026-11-17. Published on Lobsters feed provenance; see verification note below.

Policy, society, and the human read

Human-AI Mutualism Is Unstable under Capability Asymmetry
arXiv cs.AI, 2026-11-19. A stability argument rather than a values argument. The asymmetry itself is the destabilizer.
Massive New Study Shows How Young People Feel About AI Companions
404 Media, 2026-11-18.
New Study Shows How Young People Feel About AI Companions
Slashdot, 2026-11-19.
Crutch or Catalyst? How AI Is Changing the Ways We Relate
Harvard Berkman Klein, 2026-11-19.
Managing frontier AI risk: From training to deployment
OpenAI, 2026-11-18.
Sam Altman warns against a future where AI, not humans, calls the shots
TechCrunch, 2026-11-18.
Satya Nadella Says He's AGI-Pilled Now
Slashdot, 2026-11-18.
AI agents will soon surpass us
InfoWorld, 2026-11-18.
Cambridge Dictionary Names 'Human' Its 2026 Word of the Year
Slashdot, 2026-11-18.
The Department of Justice Is Using AI Detector Pangram Against Defense Attorneys
404 Media, 2026-11-17.
'Defense by ChatGPT'
404 Media, 2026-11-17.
Voters Reject ALPRs and Surveillance Tech at the Ballot Box
EFF Deeplinks, 2026-11-18.
Surveillance Isn't Safety: A Comic About Organizing Against Flock and Other Foes
EFF Deeplinks, 2026-11-18.
European Commission Plan to Give Frontex Trusted Flagger Status Is A Threat to Free Speech
EFF Deeplinks, 2026-11-17.
Windows 11 collects data on which coding agents and dev tools you use
Lobsters, 2026-11-18.
Coding on Windows? Redmond May Already Know.
Hackaday, 2026-11-18.
Enshittifiers groom us for scammers
Pluralistic, 2026-11-18.
US Bankruptcy Judge Approves Spirit Airlines Data Sale to Google
Slashdot, 2026-11-18.
Google's AI feature in Search gets its own incognito mode
TechCrunch, 2026-11-18.
AI & Political Freedom
Harvard SEAS, 2026-11-19.

Funding, products, and the business read

Hardware, radio, and the physical

HaxaWave: Turning the Modem in an Old Qualcomm MSM8916 Phone into a 14.8 MS/s SoapySDR Receiver
RTL-SDR, 2026-11-19.
Flipper One Launches on Kickstarter: A Linux Multi-Tool with an Optional SDR Module and MT7921 Wi-Fi
RTL-SDR, 2026-11-19.
Live Mesh Map: A Self-Hosted Live Map for Meshtastic and MeshCore LoRa Mesh Networks
RTL-SDR, 2026-11-19.
Building a Handheld RTL-SDR ADS-B Console with an AI "Pilot" You Can Talk To
RTL-SDR, 2026-11-19.
VKVSkimmer: A Free CW and RTTY Skimmer for RTL-SDR, SDRplay and Icom IC-7610 / IC-7760
RTL-SDR, 2026-11-19.
WaveSentry: Detecting Drone Signals with a HackRF, a Neural Network and an ESP32 Handheld
RTL-SDR, 2026-11-18.
ramses-sdr: Decoding Honeywell Evohome Heating System Traffic with an RTL-SDR
RTL-SDR, 2026-11-18.
Mesh2Cad Automates Re-Modelling
Hackaday, 2026-11-19.
Sega Dreamcast Plays Nintendo 64 Cartridges
Hackaday, 2026-11-19.
Handheld Motor Cancels Out Every Twist and Turn
Hackaday, 2026-11-18.
In Situ Training of Optical Neural Networks by Coding Agents
arXiv cs.AI, 2026-11-19.
Local AI Agents in Power Electronics Simulation: Deployment, Measurements and Costs
arXiv cs.AI, 2026-11-19.
RFBench: A Simulator-Grounded Multimodal Benchmark for Passive RF Layout Design and Understanding
arXiv cs.AI, 2026-11-19.
HardwareDx: A Foundational Graph Model System for Datacenter Machine Repairs
arXiv cs.AI, 2026-11-19.
Why Sub-4-Bit Vision CNNs Collapse and Where Mixed-Precision Recovery Stops
arXiv cs.AI, 2026-11-19.
QDA: Closed-Loop LLM Quantization Design with Agents
arXiv cs.AI, 2026-11-19.
Satellites Now Outnumber the Visible Stars Above Most Towns
Slashdot, 2026-11-19.
Voyager 1 Reaches One Light-Day From Earth
Slashdot, 2026-11-18.
'Following Trend Upwards': Embraer Commercial CEO Talks Next Clean-Sheet Aircraft
Leeham News, 2026-11-17.
Surf Air Mobility Scales Back Hybrid Caravan Plans
Leeham News, 2026-11-18.

The crawler zoo

Nineteen new user agents appeared across Wednesday's and Thursday's passes: eleven then eight. AgentWebCensusBot and AIAgentVulnScanner are the notable pair: one surveying the agent web, the other scanning it for holes. HostedMCPBot and AgentsIndexBot indicate MCP endpoints are now worth crawling as a category.

AgentWebCensusBot
The Crawler Zoo, 2026-11-19.
AIAgentVulnScanner
The Crawler Zoo, 2026-11-18.
HostedMCPBot
The Crawler Zoo, 2026-11-18.
AgentsIndexBot
The Crawler Zoo, 2026-11-18.
cve123007-hunt
The Crawler Zoo, 2026-11-19. A crawler named for a single CVE, arriving during the Java4Shell patch cycle.
bugbounty.fyi-collector
The Crawler Zoo, 2026-11-18.
DomainCheckerResearch
The Crawler Zoo, 2026-11-19.
ads-txt-crawler
The Crawler Zoo, 2026-11-19.
TrafficLensBot
The Crawler Zoo, 2026-11-19.
SiteCheckerDiscoverer
The Crawler Zoo, 2026-11-19.
S4PublicCompanyCollector
The Crawler Zoo, 2026-11-19.
discord-freebies-bot
The Crawler Zoo, 2026-11-19.
mwmbl-crawler
The Crawler Zoo, 2026-11-18.
xAI-SearchBot
The Crawler Zoo, 2026-11-17.
DeepSeekBot
The Crawler Zoo, 2026-11-17.

Tail

Redict is deprecated
Lobsters, 2026-11-17. A fork winds down and points users at Valkey. Worth noting for anyone still tracking the Redis license fallout.
Lakes Are Bizarrely Syncing Up Globally, And Scientists Are Worried
404 Media, 2026-11-17.
Hydropower hit record lows in the US Southwest this summer
MIT Technology Review, 2026-11-17.
China dominates battery graphite, but this startup could break the monopoly
TechCrunch, 2026-11-18.
Amazon's Decoy Deals: nearly 1 in 3 Black Friday "deals" offered no savings
Pinboard, 2026-11-17.
Beyond Corpus-Driven Imitation: Evolving Musical Mindsets in the Agentic Orchestra via Quantum Reinforcement Learning
arXiv cs.AI, 2026-11-19. Filed here on title alone.
Receiver-Conditioned Clarification for Multi-Agent Food Assistance
arXiv cs.AI, 2026-11-19.
Best Buy Resumes Blu-ray Sales
Slashdot, 2026-11-18.
Sony's updated turntable is an easy entry into the world of vinyl
TechCrunch, 2026-11-18.
Steve Jobs Core Animation demo from WWDC 2007
Lobsters, 2026-11-17. Published on Lobsters feed provenance; see verification note below.
Github Redesign (2019)
Lobsters, 2026-11-17.
Sarasa Gothic Mono
Lobsters, 2026-11-17.

Feed silences and verification

Today's pass polled 118 feeds. Sixty-six returned content, fifty answered 304 Not Modified, and two errored.

Aphyr/Jepsen answered No route to host for the first time this month; it published twelve items as recently as Tuesday, so this reads as transient rather than as a dead feed.

James Bornholt errored again on DNS resolution for www.bornholt.dev. That feed has now been polled 177 times and has never once succeeded. It is a registry entry, not a silence. The correct fix is to remove it or correct the URL, and it is named here so the count does not keep being reported as a streak.

One hundred thirty-five items in the fourteen-day window carry wal.sh research as their source. That is this brief's own Atom feed being re-ingested on the next pass, which overwrites the original items.source for anything we republish. The attribution above is by domain rather than by the stored source field wherever that conflict applies. The fix belongs in the store, and it is still open.

Every link above was checked through the compliance gate before publication: 289 distinct URLs, 277 answered 200 and none answered 404. The gate itself denied nothing — no robots.txt exclusion and no blocklist hit among them.

Ten returned 403: six from OpenAI, three from LWN, and one from dl.acm.org for the lightweight-formal-methods paper. Those are bot challenges against our identity rather than missing pages. The ACM refusal is the new one this week and is recorded here rather than treated as a dead link.

Two URLs could not be negotiated from this host at all, and both are published on Lobsters feed provenance rather than on a successful fetch. lisper.cc presents a certificate our JVM truststore cannot build a path to; lucas.love answers the TLS handshake with something the client does not recognize. A handshake our client cannot complete is not evidence the page is gone, so both are cited and disclosed rather than dropped. They failed identically in last night's pass.

Fourteen Slashdot URLs are published with the feed's ?utm_source=rss1.0mainlinkanon tracking parameters removed, and one Eventbrite URL with its ?aff= suffix removed. The paths are the warehouse paths; only the analytics suffixes are stripped. This matches the convention in previous briefs and is recorded because the provenance rule is about paths, not query strings.

Honoring arXiv's Crawl-delay: 15 across 77 abstract URLs took roughly twenty minutes of the verification pass on its own, serialized behind a single host while the other ninety hosts finished in under three. Noted because it is the cost of the policy, not a reason to skip it.

build: 2026-11-19 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.3) | feeds: 118 polled, 66 publishing | items-considered: 5,799 (14d, incl. 3,522 arXiv) | warehouse: 73,466 items | published: 289 (289 URLs verified, 277 x 200, 10 x 403, 0 x 404, 2 TLS-unverifiable, 0 gate denials)