feed
last scrape: 3h ago · 961 in archive · 9 today
today
YC Founder Firesides interview with Vori CEO Brandon Hill on digitizing supermarkets from clipboards/faxes (starting with inventory reordering app), family roots in grocery, early customer wins, and $22M Series B. No robots or autonomy.
Vori's AI OS Digitizes $1.5T US Grocery Retail
Explains the "agent judge layer": a separate LLM validator at the action boundary that classifies risks into four buckets and uses a four-way decision (beyond yes/no) to gate agent tools, as Lindy added after unauthorized emails. Prompts & playbook here.
Agent Judge Layer Guards Production Actions
Walkthrough of Gemini's File Search API now embedding images alongside text for cross-modal retrieval, metadata filtering (e.g., department=legal), and page-level citations — Colab demo covers upload/query flow.
Gemini File Search Adds Multimodal RAG
Outlines a five-level bash security framework for AI coding agents—user prompts (L1), system prompts (L2), blacklist hooks like Damage Control (L3), whitelists (L4), no bash at all via MCP servers or Pi extensions (L5)—demoed with destructive prompts on Claude Code and Pi Agent, including a GPT exploit via fake package.json.
5 Levels to Eliminate Bash Risk in AI Agents
NLW argues AI expands the demand frontier via six elasticities—price, access, complexity, continuity, personalization, relational—plus a lasting "human premium," using healthcare (e.g., continuous monitoring) as a case study for new work categories. Companion read.
AI Expands Economy's Demand Frontier, Creating Human-Premium Jobs
Walkthrough of Okara AI CMO, a $99/mo SaaS that analyzes your site and runs background agents for SEO audits, GEO optimization, blog writing, and Reddit/HN/X posts.
Okara AI CMO deploys site-analyzing marketing agents
Step-by-step on cross-compiling llama.cpp for ARMv6 via dockcross (no Neon/OpenMP), loading Q4/Q8 Falcon-H1-Tiny-90M on Pi OS Lite with --no-mmap and 128-token context; 4-bit coherent but slow (~3s/token), 2-bit nonsense.
90M Falcon Runs on 2014 Raspberry Pi
Reaction to Lars Faye's article on "agentic coding" causing cognitive atrophy, non-determinism complexity, skill loss, vendor lock-in, and token costs—with creator pushing back on costs while agreeing on atrophy risks. Sponsored by Browserbase.
Agentic Coding Trap: Cognitive Debt Hits Hard
yesterday
Commentary on McKinsey's Lilly platform exploit—$20 AI agent via SQL injection on 22 unauthenticated endpoints—as a procurement/strategy failure, not hygiene, plus vendor responses from Anthropic, OpenAI et al., and a 6-question checklist. Full playbook here.
Lily Hack: AI Procurement Ignores Agent Realities
Overview of Codex's new Chrome extension for signed-in browser access (Gmail, Salesforce, etc.) with allow/block lists, plus v0.129 CLI upgrades like Vim editing, plugin sharing, and hooks; v0.128 adds persisted goals and keymaps.
Codex Chrome Extension Enables Signed-In Browser Tasks
Walkthrough of enabling Codex's experimental /goal feature (add features.goals = true to config.toml) for autonomous long-running tasks like a ReAct loop with budget handling, then demoing it build a 2D survival game from plan.
Codex /goal: Simple Harness for Hour-Long AI Coding Agents
Hour-long walkthrough of installing open-source Hermes Agent on Hostinger VPS, wiring it to Telegram, adding a first skill/cron job, GitHub backup, plus its five pillars and vs. Claude Code/OpenClaw.
Build Self-Improving Hermes AI Agent on VPS
Walkthrough of Google Labs' Pomelli (free tool): input site URL to auto-extract brand/colors/tone, new Catalog pulls full product lineup, generate per-product photo shoots (e.g., model try-on templates) and full social campaigns, download ready-to-post.
Pomelli Catalog Imports Products for Scaled Campaigns
this week
Four-step audit for subtle AI hallucinations in high-stakes docs: 1) finish AI draft, 2) extract claims to table (w/ sources), 3) validate vs source via 4 labels (supported/conflicts/no-proof/needs-human), 4) rewrite—all in fresh chats. Prompts in presentation.
4-Step AI Audit Catches 'Almost Right' Errors
Live demo of Archon using "harness engineering": YAML DAG workflows, git worktrees for isolated parallel agents, and auto-loading skills with Claude Code for more consistent PRs. Model quality still matters, and workflows need upfront design.
Archon Makes AI Coding Agents Deterministic via Harness Engineering
Breakdown of Codex agent scaffolding: prompts for one-offs, skills as markdown "house styles" for reuse, plugins packaging full workflows with integrations, plus MCPs/hooks/scripts—decision tree in the article.
AI Agents Need Scaffolding: Prompts to Plugins
Walkthrough of installing vidIQ MCP connector in Claude Code to audit channel metrics (VPH trends, outliers), compare to competitors like Nate O'Brien, spot title/thumbnail gaps, and prompt a live analytics dashboard.
vidIQ MCP enables Claude to audit YouTube channels
YC Founder Firesides chat with Trigger.dev co-founders on three product pivots: v1 "Zapier for devs," v2 embedded async tasks, v3 hosted SDK execution—hitting PMF as AI agents drove 90% usage.
Trigger.dev Pivots to AI Agents, Hits PMF with 90% Usage
Walkthrough of Cursor's /goal command best practices: enable via /features enable goal, prompt with explicit verifiable "done" states (e.g., Playwright visual checks, quantified targets like "20 issues"), initial alignment chats, and scaffold via npx goalbody for goal.md + state.yaml files.
Codex /goal tips: Define verifiable 'done' states
Walkthrough and demo of Symphony, OpenAI's Elixir tool for spinning up Codex agents on Linear issues — highlights prompting an agent with a 2k-line spec to build a custom version (e.g., Python), plus hooks for repo cloning/PRs. Compares to MultiOn/Conductor but skips real-project readiness details.
OpenAI's Symphony: Autonomous Codex Agents for Linear Issues
Verdent Manager overview: coordinates AI agents by decomposing ideas into parallel tasks with long-term memory of your stack/preferences, specialist skills for testing/deployment, and Slack/Telegram messaging — invite-only, with Eco Mode/BYOK for costs.
Verdent Manager Coordinates Idea-to-Deployed App Builds
Walkthrough of Printing Press, a Go-based CLI factory/library with 50+ pre-builts (ESPN, Craigslist, etc.) that agents like Claude Code invoke token-efficiently vs MCPs (35x fewer tokens, 100% vs 72% reliability) or raw APIs. Demos setup and custom CLI for Skool.
Printing Press: CLI Factory for AI Agents
Weekly AI podcast recaps essays from Ezra Klein/Alex Ess/A16Z on relational job growth over apocalypse fears, plus Anthropic/OpenAI enterprise JVs and Wall Street data center bets as maturation signs.
AI Shifts: No Job Doom, Infra Boom Ahead
Demo of oMLX using MLX's Two-Tier KV cache to page inactive context to SSD on M2 MacBook Pro: 47 t/s and 89% cache efficiency with Qwen 3.6 (vs LM Studio's 16 t/s), enables multitasking, but hits occasional 400 context errors.
oMLX SSD KV Cache Enables 3x Faster LLMs on M2 Macs
Interview with Claude platform leads Angela Jiang and Katelyn Lesse on Managed Agents primitives (messages API, code execution sandbox, file systems, skills), production scaling hurdles, and a future of outcome+budget goals with self-writing harnesses.
Claude Managed Agents: Production-Ready AI Infra from Anthropic