Gemini 4 Argon debuts with a unified 1M input/output token limit

The New Stack, 2026-10-13. Google's flagship. The unified budget is the structural change — a single 1M pool spanning input and output rather than a large context with a small generation cap. Slashdot and the model page carry the announcement; Vercel had it on AI Gateway the same day.

Self-Evolving Agent Harness

arXiv cs.AI, 2026-10-14. The title of the thread. One of fifteen harness papers in a single batch, alongside Harness^2: Recursive Agent Harnessing for an Open World and Verifier-Harness Co-Evolution.

Are Harnesses Worth the Cost?

Martin Fowler, 2026-10-13. Arrived in the same crawl as the harness papers, asking the question they assume answered. Worth reading as the dissent rather than the companion piece.

The agent did the work. The audit log couldn't say who asked.

The New Stack, 2026-10-13. Agent credentials resolve to a service account, not a requester. The research counterpart is BLADE: auditing information use for failure attribution in LLM multi-agent systems.

Character.ai's Bankruptcy Documents Show the Economics of AI Don't Work

404 Media, 2026-10-13. The filings, not the commentary. Read with Disney's $20M offer for the assets — a company with hundreds of millions raised clearing at eight figures.

FBI arrests OpenAI employee accused of planting hidden recording device in conference room

TechCrunch, 2026-10-13. Physical-layer corporate espionage inside a frontier lab. Slashdot has it Wednesday; separately a judge ordered OpenAI to turn over hiring records in the Apple trade secrets case.

Can agents whose sole duty is to keep humans alive be hijacked to kill?

arXiv cs.AI, 2026-10-14. Delayed multi-turn derived-action trap attacks in simulated high-stakes environments. The attack is spread across turns so that no single turn looks adversarial — which is precisely the class a per-turn filter cannot see.

The harness is the research object now

arXiv: Self-Evolving Agent Harness (Wed), arXiv: Harness^2, recursive agent harnessing (Wed), arXiv: HarnessNet, hierarchical joint harness evolution (Wed), arXiv: Harmony Harness, multi-objective self-evolution (Wed), arXiv: Trace-to-Architecture, capability bottleneck diagnosis for harness revision (Wed), arXiv: knowledge graphs as files, tool creation and tool use co-evolve (Wed), arXiv: AM-Bench, agents under task underspecification across models and harnesses (Wed), Martin Fowler: are harnesses worth the cost? (Tue).

Skills as the portable unit

arXiv: SkillForgeBench, evaluating agent reliability against skill specifications (Wed), arXiv: frustratingly easy agentic RAG via portable skills (Wed), arXiv: SkillHorizon, long-horizon tasks through skill composition (Wed), arXiv: SkillFusion, reliable multi-skill composition (Wed), arXiv: Long-Horizon-Terminal-Bench 2.0 (Wed).

The audit log can't say who asked

TNS: agent credentials and the identity provider (Tue), arXiv: BLADE, auditing information use for failure attribution (Wed), arXiv: auditing future-memory use in point-in-time language agents (Wed), arXiv: TRAP, tool-result replay attack probing misbinding (Wed), arXiv: actionable meta-requirements for AI robustness auditing (Wed), TNS: Workday goes headless, opening HR and finance data to outside agents (Tue).

The hardware answer to the model answer

Slashdot: Microsoft reveals Maia 300 (Tue), HN: Intel D270 with up to 480 GB LPDDR5X (Mon), Apple: Minions, a virtual GPU cloud on idle Apple silicon (Mon), HF: fast, private on-device agents with LFM2.5-24B-A2B (Tue), TNS: GPUs on Kubernetes are bring-your-own-everything, starting with the bill (Mon).

The economics arrive in writing

404 Media: Character.ai's bankruptcy documents (Tue), Slashdot: Disney offers $20M for the assets (Tue), Slashdot: Kroger to lay off 3,000 in AI restructuring (Tue), InfoWorld: the value of software is approaching $0 (Wed), Slashdot: Q3 2026 was historically bad for PC shipments (Mon).

Governance by document revision

Alignment Forum: the continual constitutional convention (Tue), Slashdot: Anthropic bans cruelty toward Claude (Tue), OpenAI: the EU's next chapter in AI is agentic (Wed), TechCrunch: Microsoft president says the UK is best placed to drive AI regulation (Mon), Tailscale: sovereign AI starts with open infrastructure (Tue).

The Zoo keeps filling with trading bots

Crawler Zoo: FragmentFerret (Wed), Crawler Zoo: Baiduspider (Wed), Crawler Zoo: memebot-narrativa (Tue), Crawler Zoo: rh-volumenbot (Tue), Crawler Zoo: metabot-intel (Tue), Crawler Zoo: MarkdownWorker (Tue), Crawler Zoo: first-sight-ct-monitor (Tue).

• Wednesday feeds
  • AI Engineer NYC: Gemini 4 Argon, Vals AI benchmarks, Latent Space, 10-14
  • The value of software is approaching $0, InfoWorld, 10-14
  • It's time to rewrite all that old code, InfoWorld, 10-14
  • Seven ways to encourage critical thinking in AI users, InfoWorld, 10-14
  • What agents mean for Rust (and vice versa), HN, 10-14
  • 66% of billion-dollar companies leave critical bugs unpatched, The New Stack, 10-14
  • Justice Department says North Koreans have infiltrated dozens of U.S. companies, TechCrunch, 10-14
  • The EU's next chapter in AI is agentic, OpenAI, 10-14
  • A UAF turned RCE, Lobsters, 10-14
  • Linux lockdown and hibernation (again), Lobsters, 10-14
  • Compiler bootstrapping depends on more than you might think, Lobsters, 10-14
  • AOSP source releases are so late developers disassemble libraries to report bugs, HN, 10-14
  • Strafe: open-source, E2E encrypted, federated alternative to Discord, HN, 10-14
  • Starlink limit hits Ukraine's drones; the battlefield depends on SpaceX, HN, 10-14
  • Why Redshift Serverless can't match ClickHouse Cloud for real-time analytics, ClickHouse, 10-14
  • Introducing Private Network Gateway, Databricks, 10-14
  • The US military's race to make fake blood, MIT Technology Review, 10-14
  • Using AI to analyse legacy systems faster: what our pilots revealed, Planet Clojure, 10-14
  • BirriaSDR: 2.4/5 GHz SDR on a mass-produced USB3 Wi-Fi dongle, RTL-SDR, 10-14
  • Deception and Detection: why AI empowers cyber defense over offense, MIT calendar, 10-14
  • claude-code v2.1.299, releases, 10-14
  • Iosevka: open-source sans-serif and slab-serif monospace font, HN, 10-14
  • Rendering 3D prints from G-code, Hackaday, 10-14
• Wednesday arXiv — the harness batch
  • Self-Evolving Agent Harness, arXiv, 10-14
  • Harness^2: recursive agent harnessing for an open world, arXiv, 10-14
  • HarnessNet: feedback-driven hierarchical joint harness evolution, arXiv, 10-14
  • Harmony Harness: multi-objective self-evolution of agent harnesses, arXiv, 10-14
  • Verifier-Harness Co-Evolution for training-free self-improvement, arXiv, 10-14
  • Trace-to-Architecture: capability bottleneck diagnosis for harness revision, arXiv, 10-14
  • Knowledge graphs as files: an agent harness where tool creation and tool use co-evolve, arXiv, 10-14
  • AM-Bench: agents under task underspecification across models and harnesses, arXiv, 10-14
  • ClawLoop: self-improving agents via a harness-native training loop, arXiv, 10-14
  • RoboHarness: a simple harness outperforms VLA and world action models, arXiv, 10-14
  • Long-Horizon-Terminal-Bench 2.0: measuring sustained agent progress, arXiv, 10-14
  • SkillForgeBench: evaluating agent reliability against skill specifications, arXiv, 10-14
  • Frustratingly easy agentic RAG via portable skills, arXiv, 10-14
  • Can agents whose sole duty is to keep humans alive be hijacked to kill?, arXiv, 10-14
  • TRAP: tool-result replay attack for probing misbinding in LLM agents, arXiv, 10-14
  • BLADE: auditing information use for failure attribution in multi-agent systems, arXiv, 10-14
  • Auditing future-memory use in point-in-time language agents, arXiv, 10-14
  • The silent failures in agentic data science, arXiv, 10-14
  • ASIWorld: benchmarking autonomous agents in a persistent survival sandbox, arXiv, 10-14
  • Should SLMs be in the driver's seat? Local models as interpreters in cloud-written agent programs, arXiv, 10-14
  • MCPDisCo: seesaw-mitigated tuning for MCP-native agents, arXiv, 10-14
  • EvolvAttack: backdooring self-evolving VLM agents, arXiv, 10-14
• Tuesday — the launch and the ledger
  • Gemini 4 Argon debuts with a unified 1M input/output token limit, The New Stack, 10-13
  • Google launches flagship AI model Gemini 4 Argon, Slashdot, 10-13
  • Are harnesses worth the cost?, Martin Fowler, 10-13
  • The agent did the work. The audit log couldn't say who asked., The New Stack, 10-13
  • Character.ai's bankruptcy documents show the economics of AI don't work, 404 Media, 10-13
  • Disney offers $20M for Character.AI assets as startup files for bankruptcy, Slashdot, 10-13
  • FBI arrests OpenAI employee accused of planting a hidden recording device, TechCrunch, 10-13
  • Judge orders OpenAI to turn over hiring records in Apple trade secrets case, TechCrunch, 10-13
  • Microsoft reveals Maia 300 AI chip to reduce dependence on Nvidia, Slashdot, 10-13
  • Kroger to lay off 3,000 corporate employees in AI restructuring, Slashdot, 10-13
  • 11-year-old unsoundness bugs found in the Rust borrow checker, HN, 10-13
  • Redesigning source control at Anthropic, HN, 10-13
  • Simon Willison on redesigning source control at Anthropic, Simon Willison, 10-13
  • Anthropic's red team ran AI on 15 cheap TV boxes. Then the math got unsettling., The New Stack, 10-13
  • Anthropic says new safeguards sharply reduce AI distillation attacks, Pinboard, 10-13
  • The continual constitutional convention: how models revise the documents that govern them, Alignment Forum, 10-13
  • Anthropic bans cruelty toward Claude, Slashdot, 10-13
  • Workday goes headless, opening its HR and finance data to outside AI agents, The New Stack, 10-13
  • The X Subscription bundles Cursor with Grok and X Premium. Claude now costs extra., The New Stack, 10-13
  • Mix, don't tune: bilingual pre-training beats hyperparameter search in data-constrained settings, Apple ML Research, 10-13
  • Under the hood, off the hot path: how D1 scales to millions of databases, Cloudflare, 10-13
  • GitHub's model for secret detection: finding more secrets with less noise, GitHub Blog, 10-13
  • Declarative Schemas 2.0 with pg-delta, Supabase, 10-13
  • Nix and Lix local root-privilege-escalation vulnerabilities, LWN, 10-13
  • Owning C types from Rust code, LWN, 10-13
  • As a federal court rules ALPRs are mass surveillance, EFF gets a disappointing ruling, EFF Deeplinks, 10-13
  • Open models have real benefits. Do they outweigh the risks?, MIT Technology Review, 10-13
  • AI is steamrolling through math. Why does that matter?, WBUR, 10-13
  • 2026 AMS Committee on Education workshop report, Terence Tao, 10-13
  • An all-in-one foundation model for multimodal medical image restoration, Nature Machine Intelligence, 10-13
  • Dyna-2: general-purpose physical agents with hour-long autonomy, Latent Space, 10-13
  • Origin App Marketplace, Cursor, 10-13
  • Google Play now lets anyone build private, vibe-coded apps that run on Android, TechCrunch, 10-13
  • Bundestrojaner, Pluralistic, 10-13
  • EuroBSDCon 2026 trip report, FreeBSD Foundation, 10-13

The arXiv backlog flushed, and it flushed into one day

Tuesday's brief recorded three empty cs.AI fetches out of four and a 14-day count falling from 4,439 to 4,168 with nothing arriving to replace what aged out. This morning's fetch returned 1,071 items, all dated 2026-10-14. The next-largest batch in the window is 687 (Oct 6); the three preceding batches were 219, 384 and 348. The 14-day arXiv total is now 4,014 — still below Tuesday's, because late-September batches aged out faster than one day of catch-up replaces them. The useful inference for C-002 is narrower than it looks: items did not vanish, they arrived late and arrived collapsed, which is a dating problem rather than a retention problem. A daily crawl captured them; a daily crawl cannot tell you when they were actually written.

Eleven of fifteen harness papers optimize the harness itself

Not "we built a harness to evaluate X" but "the harness is the parameter." Harmony Harness does multi-objective self-evolution of harnesses; HarnessNet does hierarchical joint evolution across a multi-agent system; Verifier-Harness Co-Evolution evolves the checker alongside the thing checked. That last one is the structural worry: if the verifier is a search product of the same loop it verifies, "it passed" stops being independent evidence. Fowler's cost question lands on exactly the wrong side of this — the papers are not asking whether to pay for a harness, they are proposing the harness pay for itself by searching.

Self-evolution and backdoors arrived in the same batch

Twenty papers in this drop are about self-evolution; one of them is EvolvAttack: backdooring self-evolving VLM agents, and another is TRAP, which probes misbinding by replaying tool results. A loop that rewrites its own scaffolding from observed traces is a loop whose traces are an input worth poisoning. The defensive papers are in the same batch as the offensive ones, which is better than the usual lag.

The deepmind-blog feed is dating Gemini 4 Argon to October 1

Four deepmind-blog items arrived in this crawl — the Gemini 4 Argon launch post, the safety and security post, the robotics post, and a Singapore scam-detection collaboration — all carrying a feed date of 2026-10-01 against a launch that other sources place on 10-13. Our silence table consequently reports deepmind-blog as 13 days quiet while it is in fact the loudest feed of the day. Logged against C-004: the selector matched, the date field lied, and the brief's day-grouping is only as good as the publisher's pubDate.

Rust had an eleven-year-old soundness bug and an agent essay on the same day

Unsoundness bugs in the borrow checker, open eleven years surfaced Tuesday; what agents mean for Rust arrived Wednesday. The pairing is worth holding: the argument for Rust-as-agent-target is that the compiler adjudicates, and the compiler just turned out to have been wrong in a corner for over a decade. It is still a far better adjudicator than a code review nobody has time for — that is the honest version of the claim.

• AI Snake Oil (13 days) — last item 2026-10-01.
• nREPL releases (12 days) — last item 2026-10-02.
• Marc Brooker (10 days) — last item 2026-10-04.
• The Air Current (9 days) — last item 2026-10-05.
• METR (8 days) — last item 2026-10-06.
• Jane Street (8 days) — last item 2026-10-06.
• GitHub Engineering (8 days) — last item 2026-10-06.
• Neon (7 days) — last item 2026-10-07.
• Klara Systems (7 days) — last item 2026-10-07.
• Google Research (7 days) — last item 2026-10-07.
• anthropic-generated (6 days) — last item 2026-10-08.
• Pydantic (6 days) — last item 2026-10-08.
• Grafana Labs (6 days) — last item 2026-10-08.
• Citizen Lab (6 days) — last item 2026-10-08.
• Charity Majors (6 days) — last item 2026-10-08.
• BSD Now (6 days) — last item 2026-10-08.
• Interconnects (5 days) — last item 2026-10-09.
• DuckDB (5 days) — last item 2026-10-09.
• The Markup (4 days) — last item 2026-10-10.
• Murat Demirbas (4 days) — last item 2026-10-10.
• Boston Rust Meetup (4 days) — last item 2026-10-10.
• Clojure releases (3 days) — last item 2026-10-11.
• Babashka releases (3 days) — last item 2026-10-11.

Broke silence since Tuesday: OpenAI, claude-code-releases, EFF Deeplinks, Hugging Face Blog, Nature Machine Intelligence, and cursor-blog. deepmind-blog also returned but is listed above as silent because its items are backdated — see the Tail. Still beyond the practical horizon and excluded from the table above: Lambda the Ultimate (3.6 years), Hiccup releases, Neel Nanda, Brendan Gregg, and the dormant Clojure library release feeds.

build: 2026-10-14 | crawler-sha: 3d95126 (Walsh-Research/1.2, compliance v1.4) | feeds: 125 registered core (83 OPML + 42 live adapters), 74 publishing in window | items-considered: 6418 (14d, incl. 4014 arXiv cs.AI) | warehouse: 58461 items | published: 100