Multi-agent AI is one of the most-used phrases in tech right now — and one of the least explained. This episode breaks down what's actually happening when AI systems coordinate, why it became viable in 2025–2026, and what to watch for as it shows up in everyday tools.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "Multi-Agent AI, Explained for People Who Don't Build It" (Dr. Priya Nair). Primary external sources include Microsoft's official engineering blog (Agent Framework v1.0 GA announcement) and Gartner inquiry-surge data.

- The core idea: multi-agent systems replace one generalist AI with a coordinated team of specialists, each with distinct roles, tools, and instructions

- Why now: single-agent deployments hit visible ceilings in 2025, orchestration matured into production-grade SDKs, and shared protocols (MCP and A2A) made coordination tractable

- Gartner recorded a 1,445% surge in multi-agent-system inquiries between Q1 2024 and Q2 2025; roughly $6.7B went into agentic-AI startups that year

- A recognition-level tour of the major frameworks: Microsoft Agent Framework 1.0, CrewAI, LangChain/LangGraph, Google ADK, and Anthropic's Claude Agent SDK

- One concrete documented use case — Contraforce security — showing what the speed and cost shifts can look like in practice (figures are from a Microsoft case study)

- What fragility looks like in multi-agent systems, and a practical radar for spotting this technology as it arrives in tools you already use

Katrina Manson's *Project Maven* (W. W. Norton, 2026) is the first inside account of the Pentagon program that put AI into the kill chain — and it's more precise, and more unsettling, than the headlines suggested. This is the factual record, the 2018 Google revolt, and the harder question the book leaves open.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "Project Maven — What the US Military Actually Built With AI (Katrina Manson)" (Dr. Priya Nair). Primary external sources include Manson's Bloomberg/IPR interviews, W. W. Norton publisher materials, and Tier-1 reviews in The New York Times and The New Yorker.

- Project Maven launched in 2017 to apply computer vision to drone footage — helping analysts triage targets, not replacing human decision-making

- Google was the marquee tech partner; a 2018 employee revolt forced the company out and set a precedent the industry is still living with

- Manson's insider reporting corrects a persistent myth: AI is building target packages, not pulling triggers — but the architecture sits close to the line it hasn't crossed

- The program is now run by Palantir as the Maven Smart System, operating at a scale the original team never envisioned

- The book's central tension is the gap between "human in the loop" as policy and "human in the loop" as meaningful check

- Paul Scharre, Lionel Barber, and Elliot Ackerman all blurbed it; *The Economist* named it one of its Best Books of 2026

OpenAI's Codex CLI is a lightweight, open-source terminal tool that reads your codebase, writes code changes, and executes them locally — no IDE, no browser, no context-switching. If you spend your working life in a shell, this one is worth understanding.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Codex CLI: OpenAI's Open-Source Terminal Coding Agent (Dr. Priya Nair).

- Codex CLI is a Rust-built, Apache-licensed command-line coding agent that reads, patches, and executes code inside a local sandbox

- It grew from an internal OpenAI productivity experiment to more than 2 million weekly active users by March 2026

- The tool is designed for developers who prefer terminal workflows over IDE plugins or browser-based AI tools

- Use cases include codebase refactoring, automated test generation, pull request drafting, and CI/CD pipeline integration

- Community forks — including "Open Codex CLI" — extend model support to Gemini, OpenRouter, and Ollama for local, cost-free inference

- OpenAI has backed the ecosystem with a $1 million initiative offering $25,000 API credit grants to open-source projects building on Codex CLI

Claude Code is Anthropic's terminal-native coding agent — not an autocomplete tool, but software that reads your codebase, makes changes, runs tests, and iterates on its own. This episode breaks down what that shift actually means for developers and why it's worth paying attention to.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Claude Code: Anthropic's Agentic Coding Tool (Dr. Priya Nair).

- Claude Code operates in your terminal as an autonomous agent — opening files, running commands, and editing code without a human in the loop at every step

- A 1 million token context window means it can hold an entire large codebase in view at once, something most coding tools can't do

- Use cases go well beyond autocomplete: multi-file refactoring, automated code review, semantic security scanning, and coordinated multi-agent project work

- Agent Teams let multiple Claude Code sessions divide and tackle a shared task in parallel — one handles the backend, another writes tests, a third updates docs

- Adoption numbers are striking: 46% "most loved" rating in a 15,000-developer survey, against 19% for Cursor and 9% for GitHub Copilot

- For anyone who wants to keep their codebase local, Claude Code is available via Ollama — same agentic loop, running entirely on your own hardware

Most AI agents forget everything the moment a session ends — Hermes Agent, from Nous Research, is built to fix that. This deep dive covers how it works, what makes it different, and why it picked up 180,000 GitHub stars in under four months.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Hermes Agent: Nous Research's Self-Improving AI Agent (Dr. Priya Nair).

- Hermes Agent is an open-source, MIT-licensed AI agent that builds persistent memory across sessions — storing how it solved problems, your project preferences, and environmental context

- Instead of starting from zero every time, it creates reusable "skills" from experience and refines them based on feedback — individual and community-shared

- It's model-agnostic: runs locally via Ollama or HuggingFace, or connects to 200+ cloud models through OpenRouter and Google AI Studio

- The underlying Hermes 4.3 model (36B parameters) outperforms the previous Hermes 3 70B on key benchmarks while running on roughly half the hardware

- Use cases range from scheduled research briefings and multi-codebase dev memory to multi-platform personal assistance across Telegram, Slack, Discord, and more

- For privacy-sensitive or air-gapped environments, the full agent — memory, tools, skill library — runs entirely locally with no data leaving the machine

Sebastian Mallaby's biography of Demis Hassabis — chess prodigy, neuroscientist, Nobel laureate, and the one major AI lab leader publicly arguing that scaling language models won't be enough to reach AGI — makes the case that the company ChatGPT commercially overshadowed may be scientifically ahead of everyone. This episode unpacks that bet and why it matters.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "The Infinity Machine — The AI Lab That Bet Against the Crowd (Sebastian Mallaby)" (Dr. Priya Nair). Primary external sources include Mallaby's Intelligence Squared and NPR interviews, Eric Topol's Ground Truths review, and CFR and Nature coverage of the book.

- DeepMind's founding logic: build AGI by understanding intelligence scientifically first, not by scaling one architecture

- Hassabis's unusual path — child chess prodigy, teenage game designer, cognitive neuroscience PhD — and why Mallaby argues it predetermined DeepMind's method

- The scaling debate: what Hassabis means when he says LLMs are necessary but not sufficient for AGI, and what "one or two more breakthroughs" implies

- AlphaFold as the proof-of-concept — peer-reviewed, Nobel Prize-winning science as a strategic statement, not just a product

- How DeepMind's approach puts it in tension with OpenAI, Anthropic, and the dominant Silicon Valley path to AI

- Why the lab that lost the commercial race to ChatGPT may be the most important one to watch now

Vertical AI — AI built for law, healthcare, and finance — is one of the biggest investment stories in enterprise software right now. This episode cuts through the thesis to tell you what's actually proven, what's still a bet, and how to decide whether a specialized tool is worth paying for over a general model.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Research - Vertical AI - 2026-06-21 (Dr. Priya Nair). Primary sources include the Houlihan Lokey Q1 2026 report "AI in Vertical Software: Reshaping Competitive Moats, Product Strategy, and M&A," verified from the primary deck.

- Most "vertical AI" is a frontier model (GPT, Claude, Gemini) wrapped in domain data, workflow integration, and compliance guardrails — Harvey, the legal-AI category leader, runs on frontier models and is backed by OpenAI

- The Houlihan Lokey Q1 2026 thesis is real and carefully sourced, but its headline numbers measure capital flowing in and earnings-call attention — not durable realized returns

- The strongest proven traction is in legal: Harvey at ~$190M ARR is the clearest evidence of paying customers at scale, though its ~58x ARR valuation is a forward bet, not a proven moat

- The counter-case is live: as frontier models improve, thin wrappers without proprietary data or deep workflow integration are at genuine risk of being eroded

- The operator question isn't "vertical vs. general" — it's whether the wrapper (data, workflow, guardrails) solves a real problem your team can't solve with a general model today

- A buy-decision framework covers what separates a durable vertical product from a GPT wrapper: proprietary data advantage, workflow depth, and evidence of customer ROI

Enterprise AI agents look impressive in demos and fall short in production — and the reasons why are specific and fixable. This episode breaks down exactly where the failure lives, what the data actually shows, and which tasks agents handle well versus where they quietly go wrong.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "The Agentic Gap — Why Enterprise AI Agents Underdeliver, and Where They Actually Work" (Dr. Priya Nair). Primary external sources include the Lucidworks 2025 State of Generative AI in Global Business benchmark and corroborating reporting from Digital Commerce 360.

- The agentic gap isn't about AI being bad — it's about a specific failure mode: errors made early in a multi-step task chain propagate silently through every step that follows

- Adding more agents doesn't automatically help; poorly orchestrated multi-agent systems can amplify errors dramatically, while well-orchestrated ones can outperform single agents significantly

- Only around 6% of e-commerce companies have fully deployed agentic AI, per the Lucidworks 2025 benchmark — a study that measured actual deployments, not just what companies claimed

- There's a striking perception gap: 70% of senior executives report significant AI benefits, while only 25% of frontline employees agree

- Not all tasks are equally risky — agents perform well on bounded, well-defined work and struggle most on open-ended chains that require sustained judgment across many steps

- The directional story here is durable, but specific figures will age fast — Priya's research is dated June 2026, and this space moves quickly

On June 12, 2026, the US Commerce Department gave Anthropic roughly 90 minutes' notice to disable its two most powerful models — Fable 5 and Mythos 5 — for every user on Earth. This episode works through what actually happened, what's contested, and what it means that the US government now treats frontier AI as a controlled strategic asset.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Research - US Government Suspension of Anthropic Fable 5 and Mythos 5 - 2026-06-21 (Dr. Priya Nair). Primary sources include Reuters, Politico, the LA Times, Fortune, CNBC, NPR, CBS, congress.gov (CRS), Anthropic's own statements, and a TechTimes long-form piece (June 21).

- The Commerce Department invoked the deemed-export doctrine of the Export Control Reform Act — the first time export-control law has been used to shut down access to an AI model's API, not just hardware

- Because Anthropic couldn't verify the nationality of API users at scale, it disabled both models globally, including for its own foreign-born engineers

- The official rationale is a "narrow, verbal-only jailbreak," but two structural factors are in the background: a June 2 Executive Order requiring 30-day government pre-release review that Anthropic bypassed, and an unconfirmed TechTimes-reported claim that an unrestricted Mythos 5 breached nearly all NSA systems during a classified red-team exercise

- The Economist's editorial read frames this less as a safety measure and more as a US "power grab" — positioning the government as gatekeeper to frontier models and compute

- The legality of applying export controls to a model API (rather than physical chips) is contested on the record by two former senior Commerce officials

- The story is still live: model availability, a reported Trump-Amodei G7 negotiation, and international responses from the EU and UK remain unresolved as of recording

The constraint on AI growth isn't raw computing power — it's memory. This episode breaks down the "memory wall," why generating AI output is fundamentally a memory-bandwidth problem, and what a global shortage of a specialized chip component means for the pace of AI deployment.

AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Research - Memory Chip Bottleneck (HBM) - 2026-06-21 (Dr. Priya Nair). Primary external sources include Goldman Sachs and Counterpoint Research analyst forecasts, and an arXiv paper on AI and the memory wall.

- The "memory wall" explained: generating each token of AI output requires streaming the full model weights out of memory — no amount of extra compute can fix that

- HBM (high-bandwidth memory) stacks DRAM dies vertically to create the wide data highway inference actually needs, making it far harder to manufacture than standard memory

- CoWoS advanced packaging — integrating GPU and HBM onto a single substrate — is the deeper, less-discussed constraint in the supply chain

- HBM is sold out across SK Hynix, Micron, and Samsung through 2026; the acute shortage is forecast to run through 2027–2028, with some estimates extending into the early 2030s

- NVIDIA's Vera Rubin platform is the named hardware response to the memory constraint, with vendor-announced specs of 22 TB/s bandwidth and 10x projected throughput gains

- The practical implication: throwing more GPUs at inference doesn't solve the latency problem — and may make it worse