Rob Bailey

There are massive opportunity areas in AI adoption that no one is thinking about yet.

The cheapest insurance in your AI PM job search is overprepping for AI product sense. Most candidates I coach prep for the wrong rubric. They walk in expecting traditional product sense, the kind every PM prep guide has taught for the last 5 years: CIRCLES and templates as the framework User segments by demographics Creativity and empathy as the headline rubric Recycled answers that worked for every candidate before them That rubric scores 5-6/10 in an AI product sense round. The new rubric is different on every dimension: Model capabilities as constraints (what the model can and can't do today) Inference cost and hallucination as first-class metrics Model layer vs application layer specified for every solution Safety and probabilistic design woven into every answer The reason overprep is the right answer: the cost of prepping for the new rubric and getting an old-rubric question is maybe a weekend of extra reading. The cost of bringing the old rubric to a new-rubric round is a downlevel offer, or no offer. Asymmetric. Same 30-minute slot. Different rubric. Different outcome. Treat every loop like the new rubric until proven otherwise. Recorded the full mock with Ankit Virmani on "how would you increase Claude Code WAU 10x?" Ankit just wrapped offers at Uber, Atlassian, Cisco. The 9/10 breakdown of how he applied the new rubric is at 52:30.

AI agents are becoming a new enterprise workforce layer. 
@IrenaCronin and I write this newsletter every week.   AI agents are becoming a new enterprise workforce layer because they can move beyond simple chatbot tasks and operate inside real business workflows. They can connect with company systems, complete multi-step processes, monitor activity, and reduce routine coordination work across departments. Their value depends on strong management, clear rules, human oversight, security controls, and monitoring. Read for free and please subscribe: unaligned.io

Most people building AI agents are mixing up four distinct primitives — and it's causing architectures that break in production. Skills. MCP. Hooks. Subagents. Not the same thing. Not interchangeable. Each solves a different problem at a different layer of your agent stack. Here's the complete mental model — and why it matters now more than ever. ━━━━━━━━━━━━━━━━━━━━ THE ONE-LINE VERSION ━━━━━━━━━━━━━━━━━━━━ 𝗦𝗸𝗶𝗹𝗹𝘀 = WHAT to know 𝗠𝗖𝗣 = HOW to connect 𝗛𝗼𝗼𝗸𝘀 = WHEN to automate 𝗦𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 = WHO does the work Now let's go deeper on each one. ━━━━━━━━━━━━━━━━━━━━ ① SKILLS — The Knowledge Layer ━━━━━━━━━━━━━━━━━━━━ Skills are reusable instruction modules stored in your project and loaded on-demand by the agent. They're not always in context. That's the point. The agent reads metadata first, then loads full content only when needed. Progressive disclosure at the architecture level. Think of them like training manuals for a new hire — they don't carry the entire employee handbook in their head. They know where to look. Structure looks like this: .claude/skills/ ├── deploy/ │ └── SKILL.md ├── scripts/ └── code-review/ └── SKILL.md Simon Willison called it well: "Skills are awesome — maybe a bigger deal than MCP." I think he's right. We've been obsessing over tools and under-investing in knowledge architecture. ━━━━━━━━━━━━━━━━━━━━ ② MCP — The Connection Layer ━━━━━━━━━━━━━━━━━━━━ MCP (Model Context Protocol) is how your agent reaches the external world. GitHub. Databases. Slack. REST APIs. File systems. Internal tools. Think of MCP as the USB-C port of agentic AI. A universal interface that standardizes how agents connect to services. We're now past 10,000+ MCP servers in the ecosystem. Linux Foundation took governance in 2026. MCP Apps are becoming interactive UIs in their own right. This isn't niche infrastructure anymore. It's the connective tissue of enterprise AI. ━━━━━━━━━━━━━━━━━━━━ ③ HOOKS — The Automation Layer ━━━━━━━━━━━━━━━━━━━━ Hooks are deterministic triggers — and they run outside the agent loop entirely. Four types: ⚡ PRE-TOOL → runs before tool execution ⚡ POST-TOOL → runs after tool execution ⚡ ON-EDIT → fires when files change ⚡ ON-NOTIFICATION → alerts and logging Critical distinction: the LLM does not control Hooks. You do. That's why they're powerful for compliance, formatting, linting, and audit trails in enterprise contexts. Deterministic behavior you can actually rely on. ━━━━━━━━━━━━━━━━━━━━ ④ SUBAGENTS — The Delegation Layer ━━━━━━━━━━━━━━━━━━━━ Subagents are independent workers, not threads. Each runs in isolated context with its own model, its own permissions, its own scoped tools. The parent agent delegates — it doesn't micromanage. Real example: 🤖 Subagent 1: Code Reviewer — Tools: Read, Grep 🤖 Subagent 2: Researcher — Tools: Search, Fetch 🤖 Subagent 3: Deployer — Tools: Bash, SSH This is where genuine parallelism in agent systems comes from. And it's where most hobby-grade implementations fall short. ━━━━━━━━━━━━━━━━━━━━ THE FULL STACK ━━━━━━━━━━━━━━━━━━━━ Here's how all five layers compose together: 📦 PLUGINS package layer — bundles everything ↓ 📚 SKILLS knowledge & workflows ↓ 🔌 MCP + TOOLS external connections | built-in capabilities ↓ 🤖 SUBAGENTS isolated execution ↓ ⚡ HOOKS deterministic automation ↓ 📌 CLAUDE.md always-on context CLAUDE.md sits at the base. It's the sticky note on your monitor. Always loaded. Always in context. Project knowledge that never leaves. ━━━━━━━━━━━━━━━━━━━━ A REAL-WORLD FLOW ━━━━━━━━━━━━━━━━━━━━ Task: "Analyze competitors and write a report" Here's what actually happens under the hood: → CLAUDE.md loads project context and company info → Skill activates the 'competitive-analysis' framework → MCP fires — searches Google Drive for past briefs → Subagent 1 spawns — market-researcher gathers live data → Subagent 2 spawns — technical-analyst reviews competitor repos → Hook triggers — auto-formats output and runs linter No manual orchestration. No prompt engineering duct tape. No hoping the LLM figures it out. Just a clean, composable architecture that scales. ━━━━━━━━━━━━━━━━━━━━ THE BOTTOM LINE ━━━━━━━━━━━━━━━━━━━━ The agents that will win in production aren't the ones with the smartest LLM at the center. They're the ones with the cleanest separation of concerns between knowledge, connection, automation, and delegation. Skills. MCP. Hooks. Subagents. Get the primitives right. The rest follows. —— What part of the agentic stack are you finding hardest to get right in production?

Rocking! @gkm1 from @insightpartners explaining to WSJ reporter how he is seeing AI and agent adoption in the market.

@StockSavvyShay Do you have any adoption/usage numbers for either of these?

$NOW is expanding its $MSFT partnership by integrating AI Control Tower with Microsoft Agent 365. The goal is to give enterprises unified visibility, approval workflows and policy controls for AI agents across both the ServiceNow and Microsoft ecosystems.

As an AI Engineer. Please learn: -Prompt caching & semantic caching tradeoffs -KV cache management at scale -Speculative decoding vs quantization -RAG evaluation (RAGAS + human evals) -Cost monitoring & hidden token leaks -Agent guardrails & infinite loop detection

The Company Brain: a shared semantic state that remembers everything. Every AI tool wants to remember: meetings log calls, search stores docs, agents track tasks, workflows save actions. But siloed memory means the company still forgets.The future of enterprise AI isn’t tool-local memory. It’s the Company Brain. Read More 👇

Most people talk about Agentic AI. Very few can actually design it. Here’s a simple cheat sheet to design + explain Agentic AI architecture 👇 🎯 Start here ➡️ Define the goal What exactly should the agent achieve? 1️⃣ Orchestration Layer ➡️ The control panel Decides flow, logic, and coordination 2️⃣ Agents Layer ➡️ The workforce Single or multi-agents handling specialized tasks 3️⃣ Tools Layer ➡️ Execution power APIs, web search, databases, external systems 4️⃣ Memory ➡️ The brain Short-term + long-term context storage 5️⃣ Monitoring ➡️ The eyes Track every step, detect issues in real time 6️⃣ Reliability & Failure ➡️ The safety net Retries, fallbacks, human-in-the-loop 7️⃣ Governance & Security ➡️ The guardrails Auth, compliance, audit, data protection 💡 Real insight: Agents alone don’t make systems powerful. Architecture does. If you can explain this simply, you’re already ahead of 90% in AI. ❤️ Like 🔁 Retweet 🔖 Bookmark Follow @MeenakshiYACS for more such posts #AI #ArtificialIntelligence #GenerativeAI #CareerGrowth #Upskilling

Between 2016 and 2021, Gong grew from a tiny startup with $200k ARR, to a staggering $7.2 billion valuation and half a billion in the bank. Customers routinely told us our sales demos were 2nd to none. Here's 9 lessons I learned about SaaS sales demos I'll never forget:

@sydneyrunkle Very cool

langgraph is the runtime that powers langchain and deepagents! we've been cooking on some new features: 1. node level error handlers 2. static + dynamic node timeouts 3. delta (diff based) channels for optimized storage 4. tons of new streaming primitives our 1.2 alpha release is now out -- try it out and let us know what you think! github.com/langchain-ai/l…

“AI is only going to replace the people that don’t learn how to use AI.” Brad Birnbaum (CEO of @kustomer)

@illscience We have a way better idea. :) Coming soon…. mainwave.com.

Every ceo is being treated to Shakespearean levels of AI theatre right now and they need to know the delta between good and great. Token spend is an imperfect proxy but is useful as a disqualifying metrics (zero is bad!) and the instrumentation outlined below is a smart first step. The other qualitative measures I’m looking at: - How much are they talking about automating entire business processes vs extending individual productivity - Who at the company can push code (sales / marketing / EAs) - How has the org structure changed (collapsing sales support and operations - agents can work across these efforts) - Do you have an AI employee yet? Equilibrium is a point you pass through and at this moment CEOs must err on the side of “too much” AI because the upside is enormous and the downside of missing it is existential. Shoutout to Ann who wrote a great note on this last week: x.com/annimaniac/sta…

the reason enterprise agents aren't working isn't that the models are bad. it's that errors compound. one task at 90% accuracy is fine. three tasks chained together and you're at 73%. five and you're below 60%. that's not a model problem. that's an architecture problem. the companies that figure out how to break long-horizon work into evaluable steps and catch failures before they compound are the ones that actually automate real processes. everyone else is stuck automating small tasks and calling it AI transformation.

The Met Gala is gross. Tickets are $100k. Tables are $350k. Rich people pretend to be socialists. Vomit. It never actually mattered and now it matters even less.

@aakashgupta You missed taxes in your calculation?

Ken Griffin pulled a Harvard grad's offer. Why? Ken asked him what he'd do with $10M in the bank. "I'd quit and climb the highest peaks around the world." Ken decided that wasn't Citadel material. And the move actually makes sense. The number is not random. Citadel runs $67 billion across roughly 3,100 employees. A typical equity PM gets a ~$1.5 billion book at a 3% vol target, per published WSO economics. Hit a 1 Sharpe ratio, which is the floor for keeping the seat, and the team produces around $45 million. The team takes a 20% payout. After splits, the PM clears roughly $4 million. Two of those years puts you at $8 million. Three puts you past $10 million. That's the horizon Citadel's PM development program is built around. Three-year bonus vesting. Multi-year capital allocation. Desk costs, data infrastructure, analyst team. All of it amortizes over the seat being held long enough to compound. If $10 million breaks the candidate's utility function, they walk in year three. The option Citadel paid to underwrite has zero terminal value. The trade is upside-capped at the moment the seat finally starts producing. The mountain answer encodes the failure mode in plain English. Money is the exit. Work is the toll. Citadel's comp structure cannot transact with that utility function because the structure is designed to break it. Compare to the candidate who says "I'd buy a faster machine" or "I'd buy out my mentor's book." Same dollar amount, completely different signal. The activity is the point. Money becomes fuel for more reps. Citadel keeps that second candidate every time. A 1 Sharpe is the floor. Below it you get cut. The seat requires somebody still trying for 2 Sharpe in year four, after they personally never need to work again. That drive has to exist before the wealth. You cannot manufacture it after. Simons ran Renaissance well past his first billion. Buffett is 95 and still reads 10-Ks every morning. Druckenmiller closed Duquesne in 2010 and has traded family-office capital every year since. The candidate thought he was being hired. He was being underwritten.

ANDREJ KARPATHY COULD HAVE CHARGED $2,000 FOR THIS COURSE. He put it on YouTube. The full training stack. Tokenization. Neural network internals. Hallucinations. Tool use. Reinforcement learning. RLHF. DeepSeek. AlphaGo. 3 hours of the most comprehensive LLM education that exists anywhere at any price. Not how to use the tools. How the entire system was built from the ground up and why it behaves the way it does. The engineers who understand this build things the ones who only use the tools cannot even conceive of. The gap between those two groups is not 3 hours. It is everything those 3 hours quietly unlock for the rest of your career.

Anthropic engineer revealed exactly how they build AI agents internally. And it's not what most developers expect. Here's what stood out 👇 → Most agents fail from bad architecture than bad models → The best agents are simple. Complexity kills reliability. → Tool design matters more than prompt design → Evaluation is the hardest and most ignored part of agent development → Real production agents look nothing like demo agents

most notable events in NY this May that are worth going to: 5/4 the met gala lol 5/4 Autonomous AI HH by @Techstars 5/5 Fireside Chat with @ankurnagpal 5/6 Physical AI developer meetup 5/7 @southpkcommons design chat with @dbabbs 5/7 @M13Company x @alphaptrs investor + founder HH 5/8 Retail AI Briefing by @StartupGrind 5/11 VIP GP + Angel Breakfast at Tiffany's (by me) 5/12 @AntlerGlobal x @bklynbridgevc fireside 5/13 Cyber Founders Dinner (by me) 5/16 Founder co-working + office hours (by me) 5/23 @vercinyc spring market event links + detailed writing on NY's tech ecosystem shared via newsletter nyirl.beehiiv.com

@aakashgupta Lots of ways up the mountain. History is full of people who made declarations and then followed through.

Major cheat code in life: Stop announcing what you're going to do. Every time you tell people your plans and don't follow through, you're training them to stop believing you. The person who says nothing and delivers is worth ten people who promise everything and deliver nothing. Show up quietly.