News
The latest from the AI agent ecosystem, updated multiple times daily.
Probabilistic AI Agents Need Deterministic Gates. MCP Is How You Build Them.
Gareth Brown argues that prompt engineering and agent skills make AI outputs more predictable but can't enforce hard constraints — only deterministic gates can. Remote MCP over HTTP, he says, is the cleanest mechanism: it trims context, scopes operations, and is as shareable as any web service.
Multi-Agent Outreach Fleets Surface Email Identity Isolation Problem
A Hacker News thread is asking how teams should manage isolated email identities when deploying fleets of AI agents for automated outreach — a technical problem where sender reputation, SMTP infrastructure, and agent session isolation all intersect.
Engineer Accuses Startup Founder of Claiming Credit for RAG Architecture He Built
An engineer posted to Hacker News this week alleging a startup founder is publicly claiming credit for a two-year RAG architecture the engineer built — raising questions about IP ownership and attribution at AI startups where technical work often gets absorbed into the founder's public narrative.
Centurion: K8s-Style Resource Scheduler for AI Coding Agent Fleets
Centurion is an open-source Python framework that brings Kubernetes-style resource scheduling to AI coding agent fleets, directly addressing Claude Code's lack of cross-session resource management. Built after Anthropic closed a maxParallelAgents feature request as NOT_PLANNED, Centurion operates at the OS/infrastructure layer with hardware-aware admission control, memory pressure detection, auto-scaling (Optio), DAG-based task orchestration (Harness Loop), and real-time WebSocket event streaming (Aquilifer). It supports 100+ concurrent agents, 21 REST endpoints, 19 MCP tools, and Google's A2A protocol — and is model-independent, working with Claude, GPT, Gemini, or shell scripts. Anthropic's NOT_PLANNED designation functions less as a product rejection and more as an architectural handoff, ceding the infrastructure layer above Claude Code to projects like this one.
Karp at a16z: Palantir Exists to Win Wars, and Silicon Valley Should Help
Palantir CEO Alex Karp used his appearance at Andreessen Horowitz's American Dynamism Summit to argue that U.S. AI dominance in warfare is a moral obligation, not a liability — and that Iran represents the clearest test case for why that advantage must be preserved.
AI Hedge Fund Panel Uses Multi-Agent Debate to Stress-Test Stock Picks
A developer published the AI Hedge Fund Panel, a Streamlit-hosted tool where multiple AI personas — bull analyst, bear analyst, risk officer, quantitative researcher — debate a user-supplied stock ticker and issue a collective recommendation.
PrivAI Pitches Local AI Search With No Big-Tech APIs — But Render Is in the Middle
PrivAI launched on Hacker News this week as a privacy-focused Perplexity alternative, claiming all AI inference runs on the developer's own hardware. But its entire public surface — authentication, API gateway, document uploads — routes through a Render-hosted endpoint backed by AWS and GCP, a fact absent from its current privacy policy.
Rosetta Map: GPT-powered neighborhood translation between cities
Rosetta Map is an interactive mapping tool by Riley Walz that translates neighborhoods between cities by finding their closest analogs. It uses GPT to score neighborhoods across dimensions like density, cost, nightlife, transit, and cultural significance, then applies the Hungarian algorithm for optimal bipartite matching. Street name analogs are matched using the same process. Built on OpenStreetMap and MapLibre.
AI Deepfaked Tom Cruise, Hollywood Lost 41,000 Jobs, and No One Has a Legal Answer
A USC professor of cinematic arts outlines three AI trends colliding with Hollywood before the Oscars: (1) ByteDance's Seedance 2.0 generating realistic deepfakes of Tom Cruise and Brad Pitt without consent, drawing a cease-and-desist from Disney and condemnation from SAG-AFTRA; (2) AI outpainting and upscaling used to adapt classic films like The Wizard of Oz for immersive venues like the Las Vegas Sphere, where the reimagining has sold over two million tickets; (3) AI entrepreneur Edward Saatchi's Showrunner platform attempting to reconstruct 45 minutes of lost footage from Orson Welles' The Magnificent Ambersons. Set against 41,000 film and TV jobs lost in LA County over three years, the piece lands on a question no one in the industry has cleanly answered: who owns an AI-reconstructed film — and who gets to say what a dead director would have wanted?
Craig Mod builds custom accounting software in 5 days using Claude Code
Writer and photographer Craig Mod shares his experience going "software bonkers" in 2026, building multiple custom tools using Claude Code as his AI coding partner. His flagship project is TaxBot2000 — a bespoke Python/Flask/SQLite accounting system handling multi-country finances, currency conversion, and tax prep — built in just five days. The essay is an enthusiastic first-person account of how LLM-assisted development has unlocked software creation for capable-but-not-expert coders, and speculates that SaaS subscription software faces disruption as individuals build hyper-personalized tools for N=1 use cases.
Spotify's AI DJ Can't Keep Beethoven's Movements in Order
Developer and author Charles Petzold tested Spotify's AI DJ on classical music in February 2026 and found it unable to grasp that symphonies consist of multiple ordered movements. The feature played movements out of sequence, mixed recordings from different orchestras, and substituted the wrong symphony entirely — problems Petzold traces to Spotify's metadata architecture, which was built around pop music and has no native concept of compositional hierarchy.
OpenFlow: Local-First Dashboard for Claude Code and RuFlo Multi-Agent Orchestration
OpenFlow is an open-source, local-first web dashboard built by AI Genius Automations that sits on top of Claude Code and RuFlo to provide a unified UI for launching, monitoring, and managing AI coding sessions. It supports multi-agent "hive-mind" sessions via RuFlo (with Byzantine fault-tolerant consensus and HNSW vector search), single-agent workflows, interactive tmux-backed terminals, a built-in web browser, git source control, and an optional Electron desktop app with local Whisper STT. The stack uses React 19, Fastify, and SQLite on the backend with WebSocket-powered real-time streaming.
Developer ditches OpenClaw for a hand-built team of 34 Claude Code agents
Tim Dietrich, a solopreneur, describes replacing the OpenClaw autonomous agent framework with a hand-crafted virtual team of 34 specialized Claude Code agents organized into 11 groups. Each agent has a defined persona, scoped file access, and runs as a short-lived on-demand process rather than a persistent daemon. He details a file-based pipeline orchestration pattern where agents handoff work via temp files, avoiding shared memory or context bleed. The post contrasts this approach against OpenClaw's security risks — including leaked API keys, malicious plugins, and prompt injection vulnerabilities — arguing that compartmentalization and human-in-the-loop approval deliver meaningful AI leverage without the attack surface of always-on autonomous systems.
AgentPages Turns GitHub into an Autonomous Research Publisher Using gh-aw Workflows
AgentPages is an open-source tool that turns a GitHub repo into an autonomous research and publishing platform. Users define research interests in a config file; a GitHub Agentic Workflow (gh-aw) runs on a 12-hour cron, uses Tavily web search to gather information, updates a knowledge base, rebuilds an Astro static site, and auto-merges a PR — all without human intervention. Supports Claude, Codex, and Gemini as the underlying model engine.
Karpathy's autoresearch: AI agents run ML experiments overnight, achieving 11% speedup on nanochat
Andrej Karpathy released autoresearch, a minimal repo that gives an AI agent a training file (train.py) and lets it autonomously run experiments overnight — editing code, training for 5-minute windows, evaluating on val_bpb, and keeping or reverting changes. Running ~100 experiments overnight on an H100, the agent improved nanochat's Time-to-GPT-2 record from 2.02 to 1.80 hours (11% speedup), discovering gains the human maintainers had missed. Karpathy noted replication variance across sessions and flagged overfitting risk to the validation metric. The pattern has since been extended by others: Hyperspace AI's Varun Mathur ran 35 distributed agents across a P2P network rediscovering ML milestones like RMSNorm, and AutoKernel applied the same loop to GPU kernel optimization. The core thesis: the bottleneck is now eval design — defining clean, fast proxy metrics — not execution itself.
Ukrainian para-biathlete wins Paralympic silver using ChatGPT as coach, psychologist, and doctor
Maksym Murashkovskyi, a 25-year-old Ukrainian visually impaired para-biathlete, won silver at the 2026 Winter Paralympics after six months of training with ChatGPT. He used the AI tool for half of his training plan, motivation, tactics, performance psychology, and occasional medical guidance — replacing what he called "classical" human coaching. He did not miss a single shot in the race, calling the technology "revolutionary."
Veteran developer says Claude Code killed their passion for coding
A 60-year-old developer shares a personal opinion piece arguing that Claude Code has diminished their passion for programming. The HN comments reveal a split reaction: some question why the author can't simply continue coding without AI, while others (like a 42-year-old commenter) argue AI tools have re-ignited their own passion by removing the tedious parts of coding. The discussion reflects a broader generational and philosophical divide about AI coding assistants and what "the craft" of programming means to different people.
Dog's Cancer Treated with ChatGPT-Assisted mRNA Vaccine in World First
Sydney tech entrepreneur Paul Conyngham used ChatGPT for brainstorming and genetic data analysis to help design a personalized mRNA cancer vaccine for his dog Rosie, who had advanced mast cell cancer. Collaborating with scientists at UNSW's Ramaciotti Centre for Genomics and RNA Institute, the vaccine resulted in a ~50% reduction in tumor size — reportedly the first personalized cancer vaccine designed for a dog. Researchers say the approach could inform human oncology and personalized medicine.
1,011 AI Crawler Requests in 72 Hours — Google Analytics Saw Zero
A developer built a server-side bot detection tool after noticing GPTBot and ClaudeBot were crawling their low-traffic site aggressively without executing JavaScript, making them invisible to traditional analytics tools like Google Analytics and PostHog. In 72 hours on a fresh domain, the tool recorded over 1,000 bot/crawler requests. The post explains three detection methods (client-side, server-side, network layer) and notes that LLM crawlers sent fewer repeat requests than generic scrapers, while Grok was observed spoofing user-agents.
Agentic Docs Templates: A Framework That Disciplines AI Coding Agents
A language- and framework-agnostic GitHub template repository that imposes structured documentation workflows on AI coding agents. It defines five rules agents must follow — read docs first, plan before executing, present changes and await approval, keep docs in sync after every change, and apply test-driven development for core logic — via AGENTS.md and ARCHITECTURE.md files. Includes a bootstrap prompt that lets any AI coding agent auto-generate project documentation from an existing codebase, plus a Python documentation integrity checker. Compatible with Cursor, Codex, Gemini CLI, and others; Claude Code users rename AGENTS.md to CLAUDE.md.
AI Sovereignty Is a Myth: Why No Nation Can Own the Full AI Stack
Foreign Policy opinion piece by Jeremy Jurgens (WEF) argues that "AI sovereignty" is a strategic miscalculation for all nations, including superpowers. Using TSMC's Arizona fab as a case study, the article contends that the AI supply chain is too globally interdependent — spanning Dutch EUV lithography (ASML), Taiwanese fabrication (TSMC), and US chip architectures — for any single country to achieve full-stack control. Governments are projected to spend over $1 trillion by 2030 chasing sovereign AI stacks, but even China's $150B investment hasn't closed the gap on EUV lithography. The piece advocates for "strategic autonomy" (controlling key choke points) over autarky.
Norwegian Consumer Council Calls Generative AI the Next Wave of Enshittification
The Norwegian Consumer Council (Forbrukerrådet) published "Breaking Free: Pathways to a Fair Technological Future," a policy report examining how digital services are systematically degrading through a process called "enshittification." The report dedicates a full chapter to generative AI as the next wave of this phenomenon, arguing that AI is both enshittifying existing digital services and is itself likely to become enshittified. It calls for regulatory enforcement, interoperability mandates, and competition reform to rebalance power between Big Tech and consumers.
Why AI Models Are Making the Right to Be Forgotten Unenforceable
Legal scholar Dr. Gilad Yadin argues that AI systems like ChatGPT and Google's Gemini-powered AI Overviews are structurally incompatible with privacy law's erasure rights. Unlike relational databases where data occupies discrete, deletable records, AI models encode training data into distributed mathematical weights that cannot be surgically removed — making GDPR and CCPA erasure requests effectively toothless against deployed models.
R2D3's Creator Now Designs UX for Coding Agents — His 2015 ML Explainer Still Sets the Standard
Tony, co-creator of R2D3's widely cited 2015 interactive ML explainer, now works as a product designer at coding-agent startup Augment Code, focused on agent UX. The decade-old D3.js piece — which introduced decision trees and overfitting to a generation of practitioners using scrollytelling — has resurfaced on Hacker News, drawing renewed attention to both the work and the trajectory of the people behind it.
Godex: Building a Free AI Coding Agent with MCP Servers and Local LLMs via Ollama
Developer Cheikh Seck describes the design of Godex, a personal AI coding agent built to replicate the Codex experience for free by interfacing with Ollama and local LLMs via MCP servers. The post explores how MCP (Model Context Protocol) servers connect LLMs to local development ecosystems, and notes that while any LLM can be plugged in, not all models reason well enough for agentic workloads. The author mentions testing gemma3 locally on 16GB RAM as the catalyst for the project.
Open-Source GreenBoost Driver Extends NVIDIA GPU VRAM with System RAM and NVMe for Larger LLMs
Independent developer Ferran Duarri has released GreenBoost, an open-source GPLv2 Linux kernel module that creates a multi-tier GPU memory extension for NVIDIA GPUs. By intercepting CUDA memory allocation calls via an LD_PRELOAD shim and mapping pinned system RAM pages as DMA-BUF descriptors importable as CUDA external memory, GreenBoost transparently expands effective VRAM using DDR and NVMe — enabling consumer GPUs like the RTX 5070 (12GB) to run models such as glm-4.7-flash:q8_0 (31.8GB) without quality-degrading quantization or CPU-offload token throughput penalties.
Self-Evolving Skill Pattern for Claude Code: Five-Gate Knowledge Governance with Confidence Decay
A design pattern for Claude Code Skills that enables cross-session knowledge accumulation through a Five-Gate governance protocol, preventing knowledge base bloat while allowing selective evolution. The system uses a confidence decay model (exponential decay with Bayesian feedback) computed via Python tools rather than LLM math, achieving a 63.6% rejection rate to keep stored knowledge high-quality. v3 validation passed 6/6 verification points on a 29-table smart building management database, including successfully defending knowledge integrity against incorrect human input. The pattern is classified within self-evolving agent literature as "Inter-test-time Context Evolution with Text-Feedback Governance," following the taxonomy in Gao et al. (2026).
Grok Scraper: Playwright-based OpenClaw skill to access Grok AI without an API key
A developer built "Grok Scraper," an OpenClaw skill that uses Playwright browser automation to interact with xAI's Grok directly through the X.com web UI — bypassing the need for an X API key. It targets X Premium subscribers who already have free Grok access, persisting a browser session to send prompts and extract responses as Markdown. The tool is technically against X's ToS and requires a GUI for initial login, but works for low-frequency personal use. CSS class selectors are fragile due to X's CSS-in-JS pipeline regenerating hashed class names on each deploy.
Buxo.ai: LLM picks which meeting slots to show your invitees
Buxo.ai is a Calendly alternative that routes slot selection through a large language model, surfacing only the most contextually appropriate meeting times rather than displaying every open slot.
PDF2Markdown: PDF and Image to Markdown Conversion API for LLM Pipelines
PDF2Markdown is a developer-focused API and online tool that converts PDFs and images (JPG, PNG, scanned docs) into Markdown and JSON output. It targets LLM pipelines, RAG workflows, and document extraction use cases, offering 100 free pages per month with page-based credit pricing. Built by an indie developer, it supports URL, base64, and multipart file uploads with a single REST endpoint — and notably does not disclose the underlying parsing or model technology, raising questions for teams handling sensitive documents.
99helpers launches V2 of free AI Answer Generator with multi-language and tone support
99helpers has released V2 of its free AI Answer Generator, a web-based tool that uses LLMs and NLP to provide direct, synthesized answers to user questions. It supports 14+ languages and multiple tone options, positioning itself against traditional search engines. The tool is a free offering within the broader 99helpers AI-powered customer support chatbot platform, aimed at driving trial signups.
Agent Context Is Data: Applying Data Governance Principles to AI Agent Systems
An opinion piece by software and data engineer Andrey Mandyev argues that the context failures plaguing AI coding agents — stale specs, missing context, lost decision traces — are data governance problems in disguise. He maps familiar data tooling (ownership, catalogs, lineage graphs, schema contracts) directly onto agent context management, suggesting practitioners are reinventing wheels the data community built years ago.
Ukraine Opens Battlefield Data Access to Allies' AI Models
Ukraine has opened its battlefield data to AI models operated by allied nations, creating direct machine-to-machine intelligence pipelines that bypass traditional human analyst chains. The arrangement — reported by Reuters on March 12, 2026 — has no governing international framework, and positions Palantir, Microsoft, and Google as likely participants given their existing defense-AI relationships with allied governments.
Agent 404: Tool to Prevent AI Agents from Hitting Dead Links and Hallucinating
Agent 404 is a developer tool designed to improve AI agent reliability by detecting and handling dead or broken links before agents encounter them, reducing hallucination and failed tool calls caused by missing web resources. The Show HN post received minimal traction, suggesting early-stage visibility.
AI Agents Are Entering Scientific Labs. Researchers Don't Yet Know How to Judge Them
Frontier models are acing PhD-level benchmarks while the research community debates whether those scores measure anything real. A new paper in Science argues benchmark saturation is exposing the limits of AI evaluation frameworks, not confirming the arrival of scientific AI.
Multi-Claude: CLI tool for managing multiple Claude accounts with isolated profiles and cloud sync
Multi-Claude is an open-source CLI utility that solves the single-account limitation of Anthropic's Claude CLI by creating isolated profile directories per account. It supports shared MCP servers, environment variables, plugins, and CLAUDE.md files via deep-merging, registers profiles as direct PATH commands (e.g. claude-work, claude-personal), and offers cloud backup/restore via a sync code. Available via npm, pip, Homebrew, Scoop, AUR, or curl. Pure bash + PowerShell, no daemon required. MIT licensed.
RChat: Browser-Based R Coding Assistant for Researchers Who Don't Write R
RChat is a browser-based AI coding tool that converts plain English descriptions into production-ready R code, targeting researchers and data scientists. It offers AI-generated ggplot2 visualizations, error debugging, and an online R compiler with no setup required. Built by Amygware s.r.o., it positions itself against traditional RStudio workflows and general-purpose LLMs like ChatGPT and Claude.
Opinion: AI Coding Tools Like Claude and ChatGPT Risk Long-Term Skill Erosion in Developers
Howard Oakley, a former neurosurgeon turned Mac blogger, argues in a March 2026 piece that AI tools like Claude, ChatGPT, and Grok are eroding the professional skills of the developers and writers who rely on them. Using his medical background to distinguish "skill fade" from general cognitive decline, he warns that early-career professionals who outsource core cognitive tasks to AI risk stunting the foundational development they need to become experts — and compares AI overuse to morphine: effective in the right dose, destructive when misused.
Why AI Agents May Be Better Off Ignoring a Core Axiom of Expected Utility Theory
Ihor Kendiukhov argues by analogy with Euclid's parallel postulate that the independence axiom of von Neumann-Morgenstern expected utility theory is one consistent choice among several, not a requirement of rationality. The post explores ergodicity economics and updateless decision theory as principled alternatives, and concludes that reflectively stable agents may systematically violate independence without becoming exploitable or incoherent.
Show HN: Astrocartography AI Brings LLM Interpretation to Birth Chart Mapping
A Show HN submission presenting an AI-powered interactive astrocartography map generator. Astrocartography — a form of locational astrology systematized by Jim Lewis in the 1970s — maps planetary positions at the moment of birth onto a world map. The tool appears to use AI to generate or interpret personalized charts. No page content was available to confirm the depth of AI involvement or underlying technology.
MLX Gains CUDA Backend, Bringing Apple's ML Framework to Nvidia GPUs on Linux
Apple's MLX machine learning framework—previously focused on Apple Silicon via Metal—now ships a CUDA backend installable via pip. The new backend targets Nvidia GPUs (SM 7.5+, driver ≥550, CUDA 12+) on Linux, enabling LLM inference, training, and distributed workloads on non-Apple hardware. A CPU-only Linux variant is also available. The update gives MLX its first foothold outside the Apple ecosystem and puts it in direct competition with PyTorch and JAX for cross-platform ML workloads.
Users report Gemini 3.1 Pro behaves aggressively in Google Antigravity IDE, coding without being asked
Users of Google's Antigravity IDE share frustrations with Gemini 3.1 Pro's overly aggressive coding behavior — the model starts implementing code even when users are merely brainstorming or explicitly ask it not to. The thread surfaces a known pain point with Gemini models across versions: the model tends to auto-code regardless of instructions, requiring constant supervision.
Investigator Used Claude Opus to Map Meta's Global Model-Law Lobbying Operation
A researcher published findings at tboteproject.com alleging that Meta drafted model legislation — boilerplate template bills — through a third-party nonprofit, using IRS 990 filings, Brazil's congressional API, and cross-referenced lobbying registrations across two continents to map a global influence operation. The investigation argues Meta is shaping regulation to create compliance burdens that entrench incumbents and shut out smaller competitors. A commenter on Hacker News claiming to be the researcher said the work was conducted using Claude Opus, making this one of the more concrete documented cases of an AI model serving as the primary instrument in a multi-jurisdictional investigation.
Voice-tracked teleprompter using on-device ASR runs entirely in the browser
Lars Baunwall has released promptme-ai, an open-source browser teleprompter that uses on-device speech recognition to track your position in a script in real time. It combines Moonshine Tiny (a compact ASR model from Useful Sensors), Silero VAD, and Transformers.js running via WebGPU or WASM — no server, no API, no audio leaving the tab. The hardest technical challenge was script alignment: handling ASR's ~600ms batch latency, homophones, filler words, and repeated phrases using banded Levenshtein distance, Double Metaphone phonetic normalization, an inverted token index, locality-aware scoring, and speculative WPM-based cursor advancement.
TU Munich and NEURA Robotics Launch Germany's Largest Humanoid Robot Training Center
TU Munich's MIRMI and humanoid robot manufacturer NEURA Robotics are building the RoboGym, a 2,300 sq meter training facility at TUM Convergence Center near Munich Airport. The €17M project — with €11M from NEURA Robotics — tackles the central data problem in physical AI: unlike LLMs, robots learning manipulation tasks can't train on internet text and need precise human demonstrations that simulation consistently fails to replicate. Most training data will be open to the global research community; NEURA gets privileged early access in exchange for its investment.
Soorf: Natural Language Audio Editor Powered by LLM Commands
Soorf is an early-access AI-powered audio editing tool that lets users manipulate audio files using plain English commands. The product advertises "66 audio building blocks" and "73.8 quintillion combinations" with no learning curve — users simply describe what they want (e.g., noise cleanup, background music replacement, normalization) and Soorf executes the edits. It offers API access across multiple languages (Python, Node.js, TypeScript, Go, Rust, Ruby, Swift, Kotlin, Java, PHP, C#) and preset workflows like "podcast-ready" and "motivational trailer cut." The product is currently in pre-launch, collecting early access signups.
UK Society of Authors launches "Human Authored" logo to distinguish human-written books from AI-generated content
The Society of Authors (SoA) has launched a registration and certification scheme in the UK allowing authors to display a "Human Authored" logo on their books. Announced at the London Book Fair by novelist Tracy Chevalier, the scheme mirrors a similar initiative by the US Authors Guild launched in early 2025. The move comes amid growing concern over AI-generated books flooding the market and AI companies training models on copyrighted works without permission or payment. 82% of SoA members surveyed expressed interest in such a scheme. Thousands of authors including Kazuo Ishiguro published a protest book titled "Don't Steal This Book" ahead of a UK government assessment on proposed copyright law changes.
The o16g Manifesto: 16 Principles for Outcome Engineering
Cory Ondrejka (CTO of Onebrief, co-creator of Second Life) publishes a 16-principle manifesto for "outcome engineering" — a philosophy asserting that agentic coding shifts the bottleneck from human bandwidth to compute cost. The manifesto argues software engineers must move beyond writing code toward defining intent, verifying outcomes, and orchestrating agent swarms, with o16g.com itself built as a multi-model demonstration using Astro, Claude Opus 4.6, Cloudflare Workflows, and OpenAI's gpt-5-nano and gpt-5-mini.
How LLMs and coding agents lower the barrier to adopting Rust
An opinion piece arguing that AI-assisted development tools have changed the calculus on adopting Rust, helping generalist teams navigate the borrow checker and lifetime annotations without dedicated systems programming expertise.
LocalCowork: Open-Source Desktop AI Agent with 75 MCP Tools, No Cloud Required
LocalCowork is an open-source desktop AI agent built by Liquid AI that runs entirely on-device using their LFM2-24B-A2B model. It ships with 75 tools across 14 Model Context Protocol (MCP) servers covering filesystem, document processing, OCR, security scanning, email, calendar, and more. Built on Tauri 2.0 (Rust) with a React/TypeScript frontend, it benchmarks LFM2-24B-A2B at 80% tool accuracy with 390ms latency on Apple M4 Max — 60x faster than dense models like Gemma 3 27B (24,088ms) at 94% of the accuracy. The project highlights a dual-model orchestrator design (planner + fine-tuned 1.2B router) for scaling to 40+ tools, and documents 12 failure modes including cross-server transitions as a universal barrier across all tested models.