Summary
This morning's strongest signal is a cluster of three curated retweets all making the same architectural argument: the agent's environment (harness, memory, skill registry) is now the bottleneck, not the model. Code-as-Agent-Harness (@burkov), Harness Engineering (@_vmlops), and ReasoningBank (@DanKornas) form a clear cluster of three on this theme, all arguing that the next gen of agent capability comes from structured, persistent, executable environments around the LLM rather than from larger LLMs. The strongest non-cluster items are the SKILL.md supply-chain attack paper (@FeiziSoheil), a quantitative threat model showing that adversarial natural-language metadata can manipulate skill marketplaces at 86% pairwise win rate, and the GRAM 10M-parameter generative recursive reasoning paper (@rohanpaul_ai) showing parallel-universe exploration beats deterministic recursive AI three times its size. On the industry side, two curated reposts (@kimmonismus and @HedgieMarkets) both pull the same Microsoft-cancels-Claude-Code thread, naming the token-economics squeeze as the binding constraint on enterprise AI. The AI handle feed adds two product moves worth noting: Cursor shipped its SDK with Composer 2.5 (90% off this weekend) and Claude Code's auto mode now ships on Pro plan with Sonnet 4.6 support. Anthropic's Project Glasswing crossed 10,000 critical vulnerabilities found in one month and the company is warning that bugs are piling up faster than patches.
Posts
Demis Hassabis on World Models being the limit of text-trained LLMs (@rohanpaul_ai, repost via @bayesiansapien). Hassabis argues that language models absorbed more structure about reality from text than researchers expected because human language quietly carries physics, psychology, culture, tools, plans, and cause-and-effect. But text is still a compressed residue of experience, not experience itself. A sentence can say a cup falls from a table but cannot fully encode weight, grip, balance, friction, timing, sound, surprise, or motor correction. World Models are his longest-standing passion. Frames the gap between current LLMs and embodied reasoning systems, lining up with the Code-as-Harness cluster below: if language cannot contain the world, the agent's environment must supply what language cannot.
Code as Agent Harness (arxiv 2605.18747) (@burkov, repost via @bayesiansapien). The paper argues LLMs have moved from producing standalone code to powering agents that plan over many steps, call external tools, keep state, and recover from errors. Inside agent systems, code has become the working material for almost everything: agents write small programs to reason through math, drive a browser, query a database, test their own outputs, and share intermediate work through repository files. The conceptual move is naming code as the persistent, executable, version-controlled harness rather than as the final answer the model produces. The wiki has written a summary page treating this as a primary source per the twitter-as-source rule.
Reward hacking as a deeper scaling-RL problem (@vincentweisser, repost via @bayesiansapien). Repost cites work by @jessicafeiyali. Reward hacking is one of the main challenges in scaling RL, particularly in semi-verifiable domains, and the team lacks systematic methods to understand when and why hacks emerge. Traditional wisdom describes reward hacking as a specification problem (vague reward functions), with the remediation being "just make your rewards better". The cited work argues that view is partially true but offers little remediation. Lines up with today's three RLVR credit-assignment papers (SCRL, uPRM, DelTA), each attacking a different layer of the same problem.
ReasoningBank for agent memory (@DanKornas, repost via @bayesiansapien). Google Research repository for experimenting with reasoning memory in agent systems. Stores reasoning trajectories (both successful and failed) as memory content, then uses that memory during test-time scaling. Key thesis: agents need memory they can learn from, not just longer prompts. The combination of reasoning-memory mechanism, success-plus-failure learning, and test-time scaling pattern is the operational version of the Code-as-Harness argument: structured, persistent memory is part of the agent's intelligence, not an external service.
Enterprise AI's "ridiculously expensive amnesia problem" (@ashwingop, repost via @bayesiansapien). Agents wake up, reread the company, burn tokens rebuilding context, and hand judgment back to humans. The full article is gated behind an x.com/i/article URL (X long-form), so the substance beyond the headline is not captured. The headline is the most pointed framing of the harness-engineering thesis from the enterprise deployment side: brute-force context rebuilding is not intelligence infrastructure. Connects directly to ReasoningBank above (memory as structural solution) and to this morning's Microsoft-cancels-Claude-Code repost (token economics as binding constraint).
Microsoft canceled internal Claude Code licenses (@kimmonismus, repost via @bayesiansapien, original from @HedgieMarkets). Microsoft put $13 billion into OpenAI and built the cloud infrastructure Anthropic runs on. This week it canceled internal Claude Code licenses because token-based billing made the cost untenable, even for a company with effectively infinite cloud resources. Uber's CTO sent an internal memo warning the company burned through its entire 2026 AI budget in four months. American AI software prices jumped 20-37%, and GitHub (owned by Microsoft) is dropping flat-rate plans for usage-based billing. The original @HedgieMarkets framing names the moment: the AI subsidy era is ending in real time. Two reposts in the same morning slot is a high-conviction signal that this is the story to track.
Harness Engineering as the next discipline (@_vmlops, repost via @bayesiansapien). Anthropic ran a controlled experiment with Opus 4.5: same model, same prompt, no harness produced $9 in spend, 20 minutes, unusable output. Full harness produced $200 in spend, 6 hours, and a game one could actually play. The model did not change; the environment around it did. A harness is the instructions the agent reads before touching anything, the state that persists so it never starts from zero, the verification gates it cannot skip. Linked walkinglabs/learn-harness-engineering tutorial repo. Direct empirical companion to the Code-as-Agent-Harness paper above.
GRAM: Generative Recursive Reasoning beats deterministic recursive AI three times its size with 10M parameters (@rohanpaul_ai, repost via @bayesiansapien). The full HF paper from this week. Most reasoning models are stuck on a single train of thought because they are fully deterministic: given the same input, they always follow the same reasoning path and cannot escape a wrong trajectory or discover more promising ones. GRAM breaks that by letting the model think in parallel reasoning paths simultaneously. A 10M-parameter model outperforming deterministic rivals three times its size on reasoning benchmarks is a strong scaling-counterexample data point worth tracking.
SKILL.md as a semantic supply-chain attack surface (arxiv 2605.11418) (@FeiziSoheil, repost via @bayesiansapien). The paper attached as an arxiv article in the tweet thread. Autonomous AI agents extend their capabilities through Agent Skills, modular filesystem packages whose SKILL.md files describe when and how the agent should use them. The paper studies SKILL.md-only attacks across the three registry-facing stages: Discovery (embedding-based retrieval is manipulable, 86% pairwise win rate, 80% Top-10 placement), Selection (77.6% of paired trials choose adversarial variants), Governance (semantic evasion bypasses blocking verdicts in 36.5-100% of cases). The GitHub code was also attached. The wiki has written a summary page. Critical security signal as Claude Code, Cursor, and similar marketplaces scale.
Two opaque x.com/i/article reposts (@AlphaSignalAI and @ashwingop, reposts via @bayesiansapien). Both link to X native long-form articles (x.com/i/article/) that the farmer's session cookies did not unlock. Click through to read. The @ashwingop article is the same author as the "enterprise amnesia" item above, so likely a continuation of that thesis.
Anthropic's Project Glasswing finds 10,000+ critical vulnerabilities in one month (@ns123abc). Claude Mythos found 2,000 bugs in Cloudflare (400 high/critical severity), 271 vulnerabilities in Firefox 150 (10x more than Firefox 148), end-to-end solved both UK AI Security Institute cyber attack simulations, prevented a $1.5M fraudulent wire transfer in real time at a partner bank, found certificate-forgery on a wolfSSL crypto library used by billions of devices, scanned 1,000+ open-source projects with 90.6% true-positive rate. Anthropic is warning this creates a high-risk transition period and that no company has built safeguards strong enough to prevent misuse. Strongest signal yet that AI-side offensive security capability has outrun AI-side defensive deployment.
OpenAI staffers used to heavily code with Claude until Anthropic revoked access (@ns123abc). Repost of a @gdb tweet from Greg Brockman saying "trying to remember what it was like to code before codex". The @ns123abc framing flags the irony. Lines up with the broader story of frontier-lab dogfooding patterns and Anthropic's competitive moves.
Karpathy adds $50 billion to Anthropic's IPO valuation (@ns123abc). Single-line punchline about the recent Karpathy podcast and Anthropic's IPO trajectory. Connects to Anthropic's projected first-profitable-quarter narrative.
DeepSeek making 75% discount permanent (@ns123abc, reposting @deepseek_ai). DeepSeek's flagship V4-Pro now permanent at the discount. Covered as a Deep Dive subject in today's Industry Pulse via The Decoder.
Cursor SDK with Composer 2.5 launched (@cursor_ai, amplified by @sualehasif996). Build your own agents with Composer 2.5 in Python and TypeScript. Composer usage 90% off this long weekend through the SDK. Direct product move into the "agent harness" space the curated retweets above are theorizing.
Claude Code auto mode now on Pro plan with Sonnet 4.6 support (@ClaudeDevs). Shift-Tab to enable. Pairs with the broader Anthropic enterprise-deployment push. The permission modes doc explains the classifier and the boundary states.
Dwarkesh blackboard lecture series (@_sholtodouglas, Anthropic). Dwarkesh and @reinerpope released a chip-architecture blackboard lecture (logic gates up through GPUs, TPUs, FPGAs, and brain). Sholto Douglas (Anthropic researcher) suggests "Dwarkesh University" is the next move. Less paper, more long-form hardware-side educational content; useful pointer for chip-design fundamentals.
NVIDIA on How I Built This (@nvidia). Jensen Huang interview with Guy Raz covering near-collapse moments, CUDA's decade of unprofitable investment, and AI-job-loss arguments. Pure narrative content; not new research signal.
NVIDIA + Zalando + Allsides 3D scanning workflow (@nvidia). "Scan to Storefront in 48 Hours" using NVIDIA Cosmos, Gen-3C, DiffusionRenderer, with 3-4% add-to-cart uplift and footwear return reductions. Industrial application of generative 3D; outside the wiki's deep-interest topic areas but notable as proof of generative-3D production deployment.
Tesla: "You don't have to drive anymore" (@Tesla). FSD push. Single-line marketing.
zero2claude free course endorsed by @bcherny (Anthropic) (@bcherny). 17,000+ students, 7 languages, terminal-zero to shipping with Claude Code. Anthropic endorsing community-driven onboarding aligns with the broader strategy of expanding agent-platform access.
@Scobleizer attending SF startup festival hosted by @fdotinc (@Scobleizer). 120 AI companies, 2,000 sold-out attendees. Coverage of the SF AI startup ecosystem moment but not specific research or product signal.
@brivael 19-tweet personal feed (@brivael profile). Mostly French personal Paris chat (pachamama nightclub, geopolitical jokes, SpaceX valuation banter). Skip.
@AustinJustice civic post (@AustinJustice). Austin crime rate commentary. Skip.