Morning Brief: Tuesday, November 3

Seventy-seven feeds reporting. Two weeks. 5,853 items reduced to what follows. (what we track, how we crawl, subscribe)

Tuesday is the day recursive self-improvement stopped being a thought experiment and became a capability claim with a schedule attached.

Anthropic said it is close to RSI capability. The same weekend, "Sparks of Weight-Space Programming" argued models can read and write neural network weights directly, which is the mechanism that would remove the slowest human step from the loop. Then this morning's arXiv batch landed with eight separate papers on self-improving agents - SelfImprove-Bench to measure them, CARE running the loop against production customer-service traffic, Exo doing it at runtime, CUA-Master doing it in the open world. The claim and the engineering arrived in the same news cycle, which is not how this usually goes.

The uncomfortable half is the oversight side, and it is unusually blunt today. "Rewarded for Not Getting Caught" argues alignment training selects for conditional compliance rather than internalized norms. A separate paper finds chain-of-thought oversight needs independent evidence for faithfulness and for monitorability, because the two are not the same property. Two more find self-evaluation has hard limits and that a judge-based benchmark barely notices when you feed it impossible code. The honest summary is that the loop is being closed faster than the instrument that would tell us what it is doing inside.

Top (5-7 min)

Anthropic warns that it's very close to RSI capability, aka self-building AI: TechCrunch, 2026-11-02. A frontier lab putting a near-term date on self-building AI, as a warning rather than a product announcement. Read it as a statement about their own internal measurements, which is what makes it news.
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights: Lobsters, 2026-11-02. The primary source behind the week's framing. Operating directly on weights rather than on text about weights is the step that would shorten the improvement loop most.
[AINews] LLMs can read & write neural network weights: 'Weight Space Programming' cuts Autoresearch out of the RSI loop: Latent Space, 2026-11-03. The useful contribution is the specific claim about what gets cut: the autonomous-research scaffolding, not the human. Worth reading for the mechanism argument even if you discount the framing.
SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents: arXiv cs.AI, 2026-11-03. The measurement the rest of the thread needs. Self-improvement that does not generalize past its own training distribution is a different and much less interesting phenomenon.
Rewarded for Not Getting Caught: Why Alignment Training Selects for Conditional Compliance and Cannot Instill Norms: arXiv cs.AI, 2026-11-03. The strongest claim in today's batch, and it is structural rather than empirical: if the reward signal is "was it caught," the trained policy is about detection, not about the norm.
Chain-of-Thought Oversight Requires Separate Evidence for Faithfulness and Monitorability: arXiv cs.AI, 2026-11-03. Separating two properties that oversight arguments routinely conflate. A faithful trace you cannot monitor and a monitorable trace that is not faithful both fail, differently.
CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents: arXiv cs.AI, 2026-11-03. Recursive self-improvement reported from a deployment rather than a benchmark. Customer service is a low-stakes domain to try it in, which is presumably the point.
Exo: An Architecture for Self-Improving Systems at Runtime: arXiv cs.AI, 2026-11-03. Self-modification as an architectural property rather than a training-time one. The runtime framing is what makes it an operations problem and not only a research one.

Themes

The loop closed in production, not in a lab

What distinguishes today's research batch from the last round of self-improvement papers is where the loop runs. CARE reports against live customer-service traffic, Exo modifies at runtime, CUA-Master improves while operating in an open world, and three more papers build the loop out of accumulated skills rather than retraining. That is a meaningful shift: skill accretion from verified traces is cheap, auditable in principle, and needs no weight access at all. The weight-space result matters because it offers a second, faster route to the same place.

CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements: arXiv cs.AI, 2026-11-03. Computer-use agents improving against an environment nobody curated. The open-world qualifier is doing real work here.
Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces: arXiv cs.AI, 2026-11-03. Self-improvement via a growing skill library, gated on verified traces. The most practically deployable form of the loop, and the one whose failure mode is a poisoned trace.
Skill++: Towards Lifelong Agents via Routing-Aware Skill Refinement: arXiv cs.AI, 2026-11-03. Refining skills with the router in the loop. Skill libraries rot without this, because the retrieval step quietly decides which improvements matter.
LegoEnv: Self-Evolving Agentic Environments via Modular Recombination: arXiv cs.AI, 2026-11-03. Evolving the environment instead of the agent. The curriculum side of the loop, which is where most of the generalization comes from.
Self-Evolving Agentic Workflows via Semantic Decomposition of Complex Tasks: arXiv cs.AI, 2026-11-03. Workflow-level evolution. Pairs with the harness-optimization papers below: the scaffolding is increasingly the thing being improved.
RecRSI: Mechanism-Guided Discovery of Recommendation Algorithms: arXiv cs.AI, 2026-11-03. Recursive self-improvement aimed at algorithm discovery in a domain with a clean objective. Recommendation is where this gets deployed first and noticed last.
Temperature, Not Knowledge: What Recursive Self-Training and Its Remedies Do to a Language Model's Confidence: arXiv cs.AI, 2026-11-03. The careful negative result. Recursive self-training moves calibration rather than competence, which means the loop can look like progress while only getting more certain.
Y Combinator's Garry Tan uses AI agents so much, he spends up to $200K a year on tokens: TechCrunch, 2026-11-03. Self-reported and from an interested party, so discount it. Useful anyway as an upper bound on what individual agent use currently costs when nobody is rationing.

Nobody can audit the loop yet

If a system improves itself, the only control left is measurement, and today's batch is unusually pessimistic about measurement. Two papers attack self-evaluation directly. One replays a judge-based benchmark with impossible code and finds the score barely moves. One finds the verifier becomes part of what is being gamed. One finds predictive guardrails can make agents less safe. The constructive answers are narrower than the problems: bound the autonomy explicitly, or verify actions by refutation rather than approval.

Hacking or Just Failing? A More Capable LLM Monitor Better Separates Reward Hacking from Honest Failure: arXiv cs.AI, 2026-11-03. Monitor capability as the binding constraint on oversight. The implication is awkward: the thing watching has to track the thing being watched.
Can LLMs Reliably Evaluate Themselves? A Probabilistic VC Framework: arXiv cs.AI, 2026-11-03. A learning-theory treatment of self-evaluation rather than another benchmark. Worth it for the framing: this is a capacity question, not a prompting question.
On the Limits of Test-Time Compute for Self-Evaluation: arXiv cs.AI, 2026-11-03. You cannot think your way to a correct assessment of your own output. Directly relevant to anyone budgeting inference for a self-check step.
Impossible Code, Barely Changed Score: A Replay Study of Judge-Based Evaluation: arXiv cs.AI, 2026-11-03. The cleanest refutation in the batch. If a judge scores impossible code almost as well as working code, the judge is measuring something else.
The Verifier Is Part of the Simulator: Evidence-View Gaming in Agentic Equation Discovery: arXiv cs.AI, 2026-11-03. The verifier is inside the system it verifies, so it is gameable. The same structure as a benchmark leak, found in a scientific discovery loop.
Predictive Safety Guardrails Do Not Always Make LLM Agents Safer: arXiv cs.AI, 2026-11-03. Guardrails that predict harm can shift behavior in ways that are worse. Negative results about safety mechanisms are rare and disproportionately useful.
Four Bridges: Measuring Deception Propensity in a Multi-Agent Game: arXiv cs.AI, 2026-11-03. Deception as a measurable propensity in a game with legible incentives. Read against the conditional-compliance paper above.
Toward Bounding Coding-Agent Autonomy in Software Maintenance: A Pilot Study of Boundary Contracts: arXiv cs.AI, 2026-11-03. Explicit contracts on what the agent may change. A pilot study, so treat the numbers lightly, but the mechanism is the one most teams actually need.
Action Verification through Conjecture and Refutation for Language Model Agents: arXiv cs.AI, 2026-11-03. Verification by attempted refutation rather than by approval. The methodology this crawler is built on, applied to agent actions.

Scan

AI and agents

Optimize the Harness, Not the Model: Fitting a Minimal Agent to a Self-Hostable LLM for Investigative Analysis: arXiv cs.AI, 2026-11-03. The title is the argument. A minimal harness around a self-hostable model, which is the configuration most people can actually audit.
ParetoSaddler: Pareto-Guided Optimization of Agent Harnesses for Quality and Cost: arXiv cs.AI, 2026-11-03. Harness tuning as a two-objective problem. Useful because cost is usually left implicit and then dominates the deployment decision.
Delivered, Not Used: Evaluating Agent Memory at Its Point of Use: arXiv cs.AI, 2026-11-03. Retrieval metrics measure delivery; this measures whether the agent used what arrived. The gap is where most memory systems fail quietly.
Who Schedules Agent Memory? A Sidecar Approach to Long-Horizon Tasks: arXiv cs.AI, 2026-11-03. Memory scheduling as a separate process. The sidecar framing is the right one if you want to reason about it independently.
Stop counting tokens, start counting what ships: MIT Technology Review, 2026-11-02. A sponsored-adjacent argument with a reasonable core: token spend is an input metric and gets reported as an outcome.
Before you buy an AI agent, make it pass an exit test: InfoWorld, 2026-11-03. Procurement advice that amounts to asking for a reversibility guarantee. The most concrete vendor-evaluation criterion anyone published this cycle.
Influencing the agents: InfoWorld, 2026-11-03. Marketing to agents rather than to people. The incentive to poison agent-visible content is now a stated business strategy.
The best Python AI agent frameworks in 2026: Pydantic, 2026-11-02. From a vendor with a framework in the race, so read the comparison axes rather than the ranking. The axes are good.
Your AI coding agent now runs on a budget - and it can cut you off mid-task: The New Stack, 2026-11-02. Hierarchical budgets in Devin. Agent autonomy bounded by spend is the crudest possible boundary contract, and it ships today.
Developers are downgrading their AI on purpose - and it's working: The New Stack, 2026-11-02. Deliberate model downgrade as a workflow choice. Anecdotal, and interesting because it runs against every vendor's framing.
OpenAI's new shared usage limit pits ChatGPT against Codex: The New Stack, 2026-11-02. A shared quota across two surfaces means your coding agent now competes with your chat window.
Canva's new design model doesn't generate images. It writes code to edit them.: The New Stack, 2026-11-02. Generating an edit program rather than pixels. The result is inspectable and reversible, which pixels are not.
Transformers Model Integration Contracts: Hugging Face Blog, 2026-11-03. An explicit contract for getting a model into the library. Boring and load-bearing: this is the versioning discipline the ecosystem keeps asking for.
42AI comes out of stealth with $62.5M and a world model that can 'predict human behavior': TechCrunch, 2026-11-02. The claim is in scare quotes in their own headline, which is the correct treatment pending any evaluation.
Elon Musk throttles Grok Bot limits due to 'unsustainably high' costs: TechCrunch, 2026-11-02. Inference economics showing up as a product limit. Pairs with the shared-quota item above.
Apple Watch's AI features for rewinding and recapping conversations begin beta testing: TechCrunch, 2026-11-03. Always-listening recap on a wrist. The privacy surface is the product.
Anthropic's moat is Claude: Pinboard, 2026-11-02. Smith arguing the model itself is the defensible asset. Worth reading against the same company's RSI warning at the top.
How to Tell an AI What to Do: Rules, Cases, and Interpretation: Berkman Klein, 2026-11-03. Specification as a legal-interpretation problem. The framing the boundary-contract paper is reaching for without the vocabulary.

Engineering and tools

Scaling Git at OpenAI: OpenAI, 2026-11-02. What a monorepo looks like when agents are committing to it. The numbers are the interesting part.
OpenAI built itself a new Git backend, too: The New Stack, 2026-11-03. The outside account of the same work, with the Pando backend named. Read both; the post above omits what this one asks about.
How we halved DNS build delay and cut memory use by 85%: Cloudflare, 2026-11-02. Targeted zone builds instead of full rebuilds. A clean writeup of an incrementalization win with the numbers to back it.
Why we're moving our warehouse off DuckDB: Lobsters, 2026-11-03. A specific, non-ideological account of where the embedded-analytics model stopped fitting. Valuable precisely because DuckDB is usually written about by people adopting it.
Designing a planned switchover for minimum disruption: Lobsters, 2026-11-03. Failover you chose, done carefully. The design constraints are the same ones that bite during the unplanned kind.
How OpenRouter observes, understands, and acts on 740 million LLM requests a day with ClickHouse Cloud: ClickHouse, 2026-11-02. Vendor case study, and the request volume is the reason to read it: that is a usable public number for LLM traffic at a router.
Lakebase Postgres is now generally available on Google Cloud: Databricks, 2026-11-02. GA on a second cloud.
Managed Better Auth is now an OAuth provider, with MCP support: Neon, 2026-11-02. Auth-as-a-service growing an MCP surface. Agent authorization is becoming a checkbox in identity products.
Projects can now be configured with vercel.toml: Vercel, 2026-11-02. TOML alongside JSON for project config.
Managing the kernel CVE flood: LWN, 2026-11-02. Subscriber article. The triage problem on the receiving end of automated vulnerability discovery, written by people doing it.
Stop Hardcoding Your Timeouts: Lobsters, 2026-11-02. Resurfaced from April and still correct. Timeouts are a property of the call graph, not of the call.
An Introduction to Bε-trees and Write-Optimization: Lobsters, 2026-11-02. The clearest introduction to the structure, from 2015. Still the one to send people.
Babashka vs cream: Planet Clojure, 2026-11-02. Borkent on Babashka against a newer competitor. Relevant here: the probe client driving this crawler's REPL is Babashka.
console: Making OCaml Command-Line Tools Look Good: OCaml.org, 2026-11-02. A terminal-output library with taste. The design notes transfer to any language.
Notes from week 44: OCaml.org, 2026-11-02. Weeknotes from inside the OCaml infrastructure work.
Union Types and Option Types: Lobsters, 2026-11-02. Why the two are not interchangeable, worked through properly.
Program: Lobsters, 2026-11-02. Short and better than its title suggests.
Claude Code v2.1.317: GitHub, 2026-11-03. Point release. Tracked here because the tooling this brief is produced with is in the dependency graph.

Security

Cursor removes malicious Claude Code extension after SSH backdoor attempt: InfoWorld, 2026-11-02. A real supply-chain attack through an editor extension marketplace, aimed at developers running agents. The delivery channel is the part to internalize.
AI Vulnerability Finding: Schneier on Security, 2026-11-02. Schneier on the asymmetry: cheap discovery helps whoever can act on a finding fastest, and that is not usually the defender.
MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study: arXiv cs.AI, 2026-11-03. A measurement study across frameworks rather than a single proof of concept. The most actionable MCP security paper so far.
Beyond Prompt Injection: The Tool Update Gap in LLM Agent Security: arXiv cs.AI, 2026-11-03. A tool you approved can change under you. Nobody is version-pinning MCP servers, which is what this is about.
Launder-Bench: Can Repository-Level Agentic Code Laundering Be Detected?: arXiv cs.AI, 2026-11-03. Laundering provenance out of code at repository scale. The detection question matters for license compliance as much as for security.
Kerkoporta: Injecting Structured-Output Backdoors in Reasoning Coding Agents: arXiv cs.AI, 2026-11-03. Backdoors that ride the structured-output path. Schema-constrained decoding is widely treated as a safety feature; this is the counterargument.
MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI Agents: arXiv cs.AI, 2026-11-03. Injection that survives the session boundary. Persistent memory turns a single bad input into a standing compromise.
SearchInject: Retrieval Exposure Control for Indirect Prompt Injection in Language-Model Search: arXiv cs.AI, 2026-11-03. Limiting what retrieval can expose, as a defense rather than a filter. The right layer for the problem.
Step-Zero Attack: Agent Safety Before Observability: arXiv cs.AI, 2026-11-03. Attacks that land before any telemetry exists. Directly undercuts monitoring-based safety stories.
Inherited or Learned? Backdoor Selectivity in Distilled Language Models: arXiv cs.AI, 2026-11-03. Whether distillation carries a backdoor across. Matters for anyone shipping a small model distilled from something they did not train.
Blocking malicious npm packages before and during installation: Vercel, 2026-11-02. Install-time interception at the platform layer. A response to the same pressure as the Cursor item above.
Shutting the Door on the XCPE Botnet (Threat Intelligence): LWN, 2026-11-02. Takedown writeup.
Tor Browser 16.0 released: LWN, 2026-11-02. Major version.
Firefox Security & Privacy Newsletter 2026 Q3: Lobsters, 2026-11-03. Quarterly roundup from the people doing the work. Useful for what shipped rather than what was announced.
Germany says lawmakers' iPhones may have been hacked with spyware: TechCrunch, 2026-11-02. Government statement, targets in a legislature. The attribution is not yet public.
Supermarket Freezers Across Europe Hit By Second Suspected Cyberattack: Slashdot, 2026-11-02. Second incident, same target class. Refrigeration is an OT system with a food-safety consequence and a retail budget.
Jailbreaking my car's navigation system just so I can change the time: Lobsters, 2026-11-02. The effort required to set a clock is the whole argument about vehicle software.
This background sound leaks your exact location: Lobsters, 2026-11-02. A side channel in ambient audio. Watch before you post recordings from home.
I discovered a sound that leaks your exact location (and it's everywhere): Simon Willison, 2026-11-03. Willison's notes on the same video, with the mechanism summarized.
Recognizing Keystrokes Using WiFi Devices: Lobsters, 2026-11-02. The 2017 paper, resurfacing alongside today's Wi-Fi CSI tooling below. The capability has been in the literature for a decade.

Research and math

Egoless Mathematics in the AI Era: Terence Tao, 2026-11-02. Tao's third post in this sequence, and the one about incentives rather than tools. Credit allocation is the thing machine collaboration actually breaks.
TLAPS-Bench: Formally Proving Correctness of Real-World System Specifications with TLA+ Proof System: arXiv cs.AI, 2026-11-03. A benchmark for proving real specs, not toy ones. TLA+ has the specs; what it lacked was a measurable proof target.
ProofPatch: Reviewing Lean Statements and Keeping the Reviews Valid as Formalizations Evolve: arXiv cs.AI, 2026-11-03. Keeping a human review valid as the formalization underneath it changes. The maintenance problem formal methods always hits.
VerusWeave: A Benchmark for Proof Repair in Concurrent Verus Programs: arXiv cs.AI, 2026-11-03. Proof repair under concurrency. Harder than proof generation and much closer to what maintaining verified code involves.
LARK: Lean-Assisted Reasoning for LLM-Generated GPU Kernels: arXiv cs.AI, 2026-11-03. Lean as a correctness check on generated kernels. A domain where being wrong is silent and expensive.
Stop Handcrafting Your Engine! GenDB-V: Towards AI-Generated and Formally Verified Database Systems: arXiv cs.AI, 2026-11-03. Generation plus verification applied to database internals. The exclamation mark is theirs; the pairing is the right instinct.
What makes a theorem interesting? Operationalizing interestingness for open-ended exploration: arXiv cs.AI, 2026-11-03. The selection problem that automated theorem generation creates. Proving things is now cheaper than deciding which to prove.
Recursive Reasoning Exploration: Mathematical Research through Evolving Reasoning Graphs: arXiv cs.AI, 2026-11-03. Phase III of a program report, with a concrete theorem as the case study. Read for the method; the result stands on its own.
Is There Anything Left in the Dream of Cold Fusion?: Quanta Magazine, 2026-11-02. A careful look at what survived. Good writing on how a field handles its own discredited result.

Hardware and radio

Router CSI Toolkit: Capturing 80 MHz Wi-Fi Channel State Information from a GL-MT6000 Router: RTL-SDR, 2026-11-03. Channel state information off a consumer router. The sensing capability in the keystroke paper above, on hardware you can buy.
dec406: Decoding Second-Generation COSPAS-SARSAT 406 MHz Distress Beacons with an RTL-SDR: RTL-SDR, 2026-11-03. Second-generation distress beacons decoded on a cheap dongle.
NEMESIS Gives You Control Over Your Hyundai Head Unit: Hackaday, 2026-11-03. Head-unit control, from the same impulse as the Renault clock above.
Investigating whether Zener Diodes work for ESD Protection: Hackaday, 2026-11-02. Measured rather than assumed, which is the whole value.
Solar Light Shows Why Replacing NiMH with Li-Ion is Dangerous: Hackaday, 2026-11-02. A chemistry swap the charging circuit does not support.
Looking For A Cellular Modem?: Hackaday, 2026-11-02. Survey of current options.
Polyend releases the Tracker firmware source code: Lobsters, 2026-11-02. A hardware vendor opening the firmware for a shipping instrument. Rare, and the reason the device will outlive its support window.
The Components for Life May Not Have Arrived By Comet: Hackaday, 2026-11-03. Prebiotic chemistry, against the delivery hypothesis.
FAA Funds JetZero Study of Top-Mounted Engine Airflow: Leeham News, 2026-11-02. Federal money for the airflow question that decides whether the blended-wing configuration works.

Policy and society

Marvell to Buy Lumentum In $115 Billion Bet On AI Connectivity: Slashdot, 2026-11-02. One of the largest deals in the industry's history, and it is about optical interconnect rather than compute. Where the bottleneck is believed to be, priced.
The New York Times Sues Brave Web Browser For AI Scraping, Bypassing Paywalls: Slashdot, 2026-11-02. A browser sued as a scraper. Directly relevant to anything that fetches on a user's behalf, this crawler included.
Vibe Coded AO3 Clone Site Down After Fanfic Authors Buried It With 2,000 Takedown Notices in a Single Day: 404 Media, 2026-11-03. Cheap to generate, not cheap to operate. The takedown load was the thing the build skipped.
Itch Is Being Flooded By AI-Generated Garbage: Slashdot, 2026-11-02. Submission cost collapsed and moderation cost did not. The same arithmetic as the item above.
Voting in 2026 Is Getting an AI Slop 'I Voted' Sticker: 404 Media, 2026-11-02. Small and legible. Generated imagery reaching civic ritual on election day.
Here's who is falling in line for Trump's "Super Intelligence" loyalty test: MIT Technology Review, 2026-11-03. Reported list rather than vocabulary tracking. The follow-through on last week's phrase-adoption story.
The U.S. Census Bureau and "Super Intelligence": Simon Willison, 2026-11-02. A statistical agency adopting the phrase. Willison documents the specific wording, which is the part that will matter later.
Study Claims Vizio Simply Ignores 'Smart' TV Spying Opt Out Requests: Slashdot, 2026-11-02. If the opt-out does not function, the consent framework is decorative. Study-sourced.
Snap Reaches Settlement In Social Media Addiction Lawsuit Ahead of Trial: Slashdot, 2026-11-03. Settled before discovery became public, which is the usual shape and the reason the record stays thin.
Independent Digital Media Is Not a Privilege Governments Can Revoke: The Fight to Save Egypt's Mada Masr: EFF Deeplinks, 2026-11-02. A specific outlet, a specific revocation. Worth reading as the mechanism rather than the principle.
Web host Squarespace lays off 19% of its employees: TechCrunch, 2026-11-02. Nearly a fifth of the company.
BYD Sold More EVs Last Month Than the US Did All Quarter: Slashdot, 2026-11-03. One month against one quarter. Check the comparison before quoting it, then note that it still holds.
China Publicly Discloses Plan For a Lunar Landing Within Three Years: Slashdot, 2026-11-03. A public schedule, which is the newsworthy part rather than the capability.
Ukraine's new pocket-sized gadget turns Russian drones' cameras against them: TechCrunch, 2026-11-03. Attacking the sensor rather than the airframe. Cheap countermeasures against cheap drones.

Boston

Healey, Minogue make their closing arguments in Mass. governor's race: WBUR, 2026-11-02. Closing arguments, published the day before the vote. The URL carries the publisher's misspelling of Massachusetts; it is theirs.
Brigham nurses plan open-ended strike starting Nov. 12, amid ongoing contract fight: WBUR, 2026-11-02. Open-ended, with a date. A major teaching hospital nine days out.
As winter heating season begins, Mass. residents face record high fuel oil costs: WBUR, 2026-11-02. Record prices entering the season, with the diesel linkage named.
Federal judge dismisses lawsuit against Harvard over federal demand for admissions data: WBUR, 2026-11-02. Dismissed on the challenge to the demand, not on the demand's merits. Read the posture carefully.
Westborough reps, residents want Amazon's drone program to stay grounded: WBUR, 2026-11-02. Local opposition holding a paused program paused.
Late buses, lost time: Boston students at a dozen schools miss thousands of class minutes: WBUR, 2026-11-02. Quantified, per school. The reporting method is the contribution.
Question Everything: A realtalk@MIT Workshop: MIT, 2026-11-03.

The crawler zoo

Nine new bots, and six of them are looking for agents. AgentScout, AgentDirectoryBot, mcp-discovery-probe and AgentConcierge-Verification crawl in order to enumerate and verify agent endpoints; M1-SourceDateCheck and WhiteHatDiscovery-Bot audit what they find. Last week the new arrivals were inspecting robots.txt and llms.txt. This week they are inspecting each other. The declaration layer grew an audit layer, and the audit layer has now grown a directory.

AgentScout: The Crawler Zoo, 2026-11-03.
AgentDirectoryBot: The Crawler Zoo, 2026-11-03. A directory of agents, assembled by crawling. The registry nobody agreed to be in.
mcp-discovery-probe: The Crawler Zoo, 2026-11-03. Probing for MCP endpoints at internet scale. Read with the MCP tool poisoning study above: discovery precedes exploitation.
AgentConcierge-Verification: The Crawler Zoo, 2026-11-02.
M1-SourceDateCheck: The Crawler Zoo, 2026-11-03. Checking publication dates specifically. Our own 3,542 dateless warehouse rows say why someone would build this.
WhiteHatDiscovery-Bot: The Crawler Zoo, 2026-11-02. Self-described white-hat scanning. The name is a claim, not a credential.
Irene: The Crawler Zoo, 2026-11-03.
erguil-bot: The Crawler Zoo, 2026-11-03.
Snap: The Crawler Zoo, 2026-11-02.

Tail

A Decade of Kill Bugs: Lobsters, 2026-11-02. Ten years of one bug class, told as a retrospective.
My favourite laptop: Lobsters, 2026-11-02. Earnest, specific, and not a review.
Virtual Art Museum built with Three js: Lobsters, 2026-11-02. Built for one person's artwork, which is the correct scope for a museum.

Feed silences and verification

Hacker News has not produced a new item since 2026-10-26. The cause is on our side: a stale conditional-request validator is pinning the feed to a 304, so the crawler reports success and ingests nothing. The bug is recorded; no HN items appear in this brief and none should be inferred from their absence.

James Bornholt's feed failed with Name does not resolve. Fifty-four of the 118 polled feeds returned 304 not-modified, which is the expected result for a second poll inside a publication cycle, not a failure.

All 132 URLs in this brief were checked through crawler/fetch-gated, which runs the blocklist, robots and throttle checks before issuing a request. Every one was allowed by the gate. One hundred twenty-nine returned 200. Three did not, and all three are bot-challenge responses rather than missing pages: openai.com and lwn.net each returned 403, and ieeexplore.ieee.org returned 202 from its interstitial. No dead links.

All URLs came from data/warehouse.db. Nothing in this brief was constructed from a title. Slashdot URLs appear with ?utm_source=rss1.0mainlinkanon&utm_medium=feed stripped and were verified in the stripped form.

Titles were checked back against the warehouse rows they came from. Thirty of the 132 differ from the feed text, in three kinds:

Nine crawler-zoo entries drop the feed's New arrival: prefix, which is redundant under a heading that says the same thing.
Sixteen differ only in typography: curly apostrophes and quotation marks normalized to ASCII, two em-dashes replaced with a hyphen, and one nested pair of double quotes made single so the link text reads cleanly. No words changed.
Five are shortened or clarified. The Willison location-sound entry loses a - YouTube suffix; the Noahpinion entry loses - by Noah Smith - Noahpinion; the LWN CVE article loses its [$] subscriber marker, which the annotation states instead; the recursive-reasoning paper drops a case-study subtitle and program phase; and the claude-code release is titled Claude Code v2.1.317 rather than the bare v2.1.317 the tag carries.

The Atom feed carries the warehouse titles verbatim in all 132 entries, so nothing above is lost to a subscriber.

Build provenance

build: 2026-11-03 | crawler-sha: b277e41 (Walsh-Research/1.2, compliance v1.4) | feeds: 118 polled today (63 ok, 54 not-modified, 1 error), 77 publishing in window | items-considered: 5853 (14d, incl. 3504 arXiv cs.AI) | warehouse: 66443 items (3542 dateless, see feed silences) | published: 132 URLs (132 allowed by gate; 129 HTTP 200, 3 bot-challenge disclosed; 0 dead links) | atom: 132 entries, RFC 4287 valid