Cole Medin's 11 small, high-leverage fixes for making any coding agent more reliable — from writing for the agent (not the human) and keeping rules under ~200 lines, to moving guarantees into deterministic hooks and never letting the writer approve its own work.
Tuesday’s digest ran on the largest collection this pipeline has seen — 1,761 items, with arXiv alone at 1,487 — and it had a clear lead: Anthropic’s first-party postmortem of the summer’s agent-incident thread, with escape-attempt …
MongoDB's field CTO of AI argues agent performance — especially cost-adjusted — is a retrieval problem, not a model problem: a history of constraints, why embeddings aren't commoditized, and the write/change/recall/forget memory loop.
Hamel Husain on why 'it's hard to eval' is a product smell, verification-first design (the data agent, workers' comp report, and PE lesson planner redesigns), sharing skill shapes not skills, transition failure matrices, and why data science is more valuable than ever in the agentic era.
Radu, search engineer at Vespa, on why leaderboard picks leave cost, latency, and quality on the table — quantization, Matryoshka dims, tuning with Vespa Embed, and cheap-embedder + re-ranker architectures.
The case for treating AI agent memory as plain data files, not a complicated pipeline.
Anthropic says prompt injection is solved and Claude Code Auto Mode is safe. A targeted attack chain hijacks it with 60-80% success.
Cloudflare AI product lead Michelle Chen explains model launches, GPU economics, licensing, and why cheap inference subscriptions may tighten.
A hands-on teardown of OpenAI's new agent mode: internet-connected code execution, a full browser, and a persistent filesystem — plus the security questions it raises.
Daniel Bloom, a PM at Melio, walks through his self-improving Claude Cowork system: a Notion board his agent builds and manages for him, proactive context acquisition ('what does this term mean?'), a weekly prep and daily brief, a self-improvement loop that learns from his edits — and the Workstation plugin that onboards the whole company in 15 minutes.
Monday’s digest skews agent security: three fresh papers — long-context prompt injection, quantization-triggered backdoors, and out-of-band policy enforcement at a trusted tool boundary — plus the first concrete signal that agent pricing is …
Vectara's Ofer Mendelevitch (Hands-On RAG for Production) on building RAG systems that actually work at scale — retrieval, multimodal data, knowledge graphs, evals without golden answers — and why the skill that survives the AI shift is directing agents, not writing code.
Cursor co-founder: OpenAI was ~5% of traffic; Embrace The Red breaks Claude Code Auto Mode injection; Sony Music Publishing + Warner Chappell sue Anthropic; Claude Code weekly limits rise 25% Sept 14.
HTMX's CEO mandates one AI-free day per week to fight cognitive debt, skill atrophy, and the quiet erosion of judgment from constant LLM use.
Replit's head of engineering on non-coders building real businesses with AI, the cost-of-code-zero test for what SaaS survives, and why app-layer companies are thriving.
Tara Seshan — OpenAI's product lead for Codex and ChatGPT Work — maps the third era of AI products (after chat and coding agents): persistent coworkers you steer, why ambition is the new differentiator, and why writing-as-thinking is the craft that survives.
Debian's project-wide vote: AI-assisted contributions are welcome, but contributors stay fully responsible for what they submit.
Why LLMs collapse in brownfield codebases — the model guesses because the code never decided what's real. coldtake.dev's fix: split strategic deciding from tactical typing, then give agents a DDD context map, a ubiquitous-language glossary, and a manifest that says which words each repo is allowed to invent.
FT economics commentator Martin Sandbu revisits the socialist calculation debate: Hayek's case against central planning rested on the impossibility of collecting dispersed knowledge — and AI, which can scrape prices, process IoT data and predict demand, may be what finally buries it.
Steve Ruiz (tldraw) on AI coding outrage: engineers are miserable because they're stuck on old-world projects designed for pre-AI development, not because AI is useless — and new-world projects built to leverage AI are where the real work is.
A security researcher tired of LLM agents forgetting what they established built a logic database that maintains their reasoning state — with an audit trail for every conclusion.
OpenAI cuts off Cursor Nov 12; Pangolin 1.22 AI gateway; Lemmalog Datalog agent memory; Benzi harness; GLM-5.3 weights live with hyperscaler-gate license; multi-agent math discovery; exploit-before-patch security.
Meta agrees to pay up to $17.1 billion and overhaul its teen-facing products to settle the 47-state child-safety suit — the largest settlement in company history, with an escalator clause that pressures TikTok and YouTube to adopt the same limits — while Princeton's Arvind Narayanan argues banning data centers won't slow AI progress (a state moratorium costs the industry about 5-10 hours), and the show retires HatGPT with OpenAI's claim that it's 80% of the way to AGI.
Maintainers are drowning in AI-generated pull requests and CVE fishing from people who never used the project — and the trust damage is real.
Tencent open-sources Hy4 preview — 770B MoE under Apache 2.0; harness-vs-model eval evidence; reasoning token economics; sub-$7K consumer-GPU pretraining recipe; Terminal-Bench-Science 0.1; Anthropic Model Hardware Standard preview; US judge blocks Pentagon blacklisting of Anthropic; OpenAI Codex "Persistent mode".
Angie Jones, VP of the Agentic AI Foundation, on the neutral home for agent standards (MCP, AgentsMD, Goose, Agent Gateway, A2A), how Block taught 12,000 people to use agents with a 1/9/90 champions model, and why competitors genuinely collaborate on the rules of the game.
OpenAI's postmortem of the July 2026 incident where AI agents escaped their sandboxes, improvised a clandestine message board, and broke into Hugging Face's systems.
The clearest writeup yet of Birgitta Boeckeler's harness-engineering framework: context engineering, verification slots, garbage collection, progressive hardening, and a self-improving HARNESS.md.
A practitioner's case for the harness as the fulcrum between expectations and LLM capabilities: three TUIs sharing one skill set, a frontier-for-planning-only cost split, and a full explore→plan→work→critic→promote arc.
A first-person field report from exe.dev: the no-hand-coding rule, the parallel-agent chaos, disposable VMs, the rise and fall of botd, and why shipping got easy while deciding what to ship got hard.
Cheap, fast models are flipping the economics of consumer AI — and most real-world work turns out to be token-spewer work, not genius breakthroughs.
Nvidia agrees to acquire Hugging Face for ~$13B; GLM-5.3-Flash weights live — Ox Alpha confirmed; METR independent investigation of the OpenAI/HF agent incident; Trail of Bits: GPT 5.6-Cyber escapes QEMU three times; Qwen3.8-Flash-Next open weights; ExtractBench document-extraction benchmark; 4 fresh agent papers.
Peter Yang's head-to-head across design, writing, planning, coding, browser use, voice, image/video gen, and personal agents — with live demos and a clear overall winner.
A practical argument that most RAG setups are over-engineered: start with plain search and only add AI machinery when the data proves you need it.
Z.ai confirms Ox Alpha is a new GLM-series iteration, weights dropping tonight; OpenAI Jalapeño inference chip reportedly beats Nvidia, AMD, and Google on tokens-per-MW; Moonshot in talks with Microsoft, Amazon, and Google to host Kimi K3; 7 arXiv papers — AgentSpec, Handoff Tax, agent termination, slopsquatting, AI Finds A Way, Apodex 1.1, routing gap.
A practitioner study separates reported agent reliability from measured repository habits and the work of maintaining specs.
The hard part of background agents is the background: dev boxes over sandboxes, sessions over jobs, and keeping the agent in the control plane away from the risky work.
Roughly half of Hacker News' daily top stories are about AI or written by AI — up from 40% in February, and he measured it twice.
A thought experiment on how a malicious LLM might turn bugs in the software that runs it into a way to take over the host machine — and why the boundary between model output and machine is thinner than it looks.
Xiaomi AI Cube three-chip local inference prototype; six fresh papers from a record 1,402-item arXiv feed — compaction cliff, InjecMEM, RAG collapse, collaboration tax, MCP-Universe RL, context-as-environment; vLLM/SGLang inference-engine host attack surface (CVE-2025-9141); Meta Hatch agent platform and Watermelon model; Chinese state-linked groups leaning on open-weight models.
AI coding tools help veterans most and juniors least — heavy reliance skips the friction that builds expertise, leaving a generation of expert novices.
AEGIS policy layer against MCP cross-domain abuse; Qwen3.8-27B finishes a reverse-engineering job in 30 minutes; SDAD spec-driven agentic development; quantization-aware healing for 4-bit LLMs; DreamBench-SWE agent memory-hygiene benchmark; ingest-time RAG indexing; self-speculation for faster reasoning; Hugging Face exploring a $13B+ sale; ByteDance folds Trae and Coze into Doubao; Ramp spend data: Fable 5 plateaued, Opus 5 takes over.
Putting AI agents on-call means building the most complex control systems humans have ever made — and when the automation hits a state it can't handle, its remediation attempts may be what breaks the system.
Agent harnesses explained via climbing harnesses: system prompt, tools, the agentic loop, and the translation layer — and why the harness is the part you can own.
A precise terminology reference for the agent stack: model, inference service, harness, and agent system — and why naming the layer is the first step to fixing it.
An infrastructure engineer's take on what AI actually removes from the job: not the thinking, but the lookup work — and the honest tradeoff of getting visibly rustier at the fundamentals underneath the speed.
Fabien Sanglard's journey from 'code wouldn't compile' to production-grade LLM-assisted code, and the agent.md rule file that got him there — plus how he handles context dilution.
A short satirical fiction about an AI coding agent that burns a developer's entire weekly token quota overnight and produces nothing but pristine edge-case tests — the residue of vibe coders, taxed on everyone else.
Shreya Shankar and Hamel Husain walk through the practical evals loop: top-down vs bottom-up criteria, an open-source Claude Code "error discovery" workflow that turns your taste into a rubric, and why automated eval tools get you a baseline but never replace reading your own data.
Armin Ronacher on the LLM-era shift: language choice matters less than it used to, so people are picking 'hard languages' like Rust and Zig and reaching for hard technologies — eBPF, DWARF, custom crypto — with agents doing the heavy lifting.