Mar 31, 2026
Erik's Newsletter: The agents are taking over
Issue #89
Today at a Glance:
Quote: Eric Hoffer on learners vs. the learned
Thinking: Agent Resources departments, services as the new software, autoresearch for everything, company knowledge graphs, the attention tax
Podcast: The design process is dead, negotiation tactics, inside Ramp’s AI-native org
Tactics: Familiar, Proof, Fort
Try: Clear your allostatic load, audit your meetings as products
Tweet: PM role splitting, infinite canvas, curiosity over capability
#QUOTE
Quote I’m pondering...
“In a world of change, the learners shall inherit the earth, while the learned shall find themselves perfectly equipped to deal with a world that no longer exists.” -- Eric Hoffer
Every week I talk to people who are “learned.” They’ve mastered a process, a tool, a role definition. And every week those definitions get less stable.
The PM who writes specs for engineers is learned. The PM who writes specs for agents, tests them, ships them, and iterates in the same afternoon is learning. The designer who runs five-stage discovery is learned. The designer who goes from prompt to prototype in a day and spends the rest of the week on taste is learning.
The gap is willingness to throw out what worked yesterday.
#THOUGHTS
What I’m thinking about...
Agent Resources: The New HR
Companies will need Agent Resources departments. AR, not HR. Someone responsible for deploying, onboarding, coaching, and quality-checking agents the way we currently do for humans.
Ramp is already operating this way. Their CPO Geoff Charles: “If you’re not using Claude Code this year, no matter what your role is, you’re probably underperforming.” Not just engineers. Sales, account managers, operators, designers. 50% of Ramp’s code is AI-written. PMs ship production features, not prototypes. The spec is the product.
Dan Shipper’s “pirate and architect” model maps here. The pirate owns a product end-to-end and ships fast. The architect turns it into a reliable system. You need a pirate for every product. You often only need an architect after PMF, and not full-time. Two humans. Hundreds of agents. AR manages the hundreds.
Services Are the New Software
The largest winners will look like services businesses selling outcomes, not tools. Every model improvement makes the work cheaper. Pure tools risk becoming a model feature.
Intelligence is automatable. Judgment is not. Copilots sell the tool. Autopilots sell the work. If you’re building a copilot, you’re competing with the next model release. If you’re building an autopilot, you’re competing with incumbents who do the work badly and expensively.
Headless services designed for agents will be the infrastructure layer. Email-for-your-agent. Calendar-for-your-agent. The companies that win will be the ones agents choose to work with.
Simulate Before You Build
Karpathy released autoresearch: an autonomous loop that makes small changes, runs experiments, scores against a concrete metric, and keeps or reverts. If you can define a clear metric and constrain the search space, you can run this loop on strategy docs, product specs, or sales playbooks.
https://github.com/karpathy/autoresearch
The Attention Tax
Human attention is degrading at the same pace AI tools improve. Anna Lembke’s framing: we are the subjects and the casualties of the largest uncontrolled experiment in human psychology ever conducted.
The attention economy is a marketplace where your consciousness is the product. Trillion-dollar industries compete for every second of your focus, using decades of behavioral psychology to exploit every vulnerability. You wake up after eight hours and still want to stay in bed. You consume hours of content and can’t recall a single thing worth remembering.
The people who win with agents won’t have the best tools. They’ll be the ones who can still focus long enough to direct them. Attention is the raw material. If yours is depleted, no amount of agent infrastructure saves you.
#PODCASTS
My favorite listens...
The Design Process Is Dead (Jenny Wen, Lenny’s Podcast)
Jenny left a director role at Figma to return to IC at Anthropic. Going backwards on the org chart to go forward on the capability curve. Traditional discovery-mock-iterate is dying because AI can mock faster than you can think. What remains is taste and the ability to evaluate. She breaks design work into two new types, and neither looks like what most design teams do today.
The Negotiation Playbook (Jacob Warwick, Lenny’s Podcast)
“What’s the chance there’s a little more here?” The most underused sentence in compensation. Warwick sees 40% average movement when negotiations are run well. Treat your job search like enterprise sales. You’re positioning, not begging. Most people lose money because they anchor on current comp instead of the pain they solve.
Inside Ramp: Where AI Agents Run Everything (Geoff Charles, Peter Yang)
Ramp’s voice-of-customer agent does 8 days of research in 8 minutes. PMs ship production code, not decks. Ramp built an L0-L3 framework to get every employee building with AI, from “use it for one task” to “build production features.” I think every company will need a version of this within 12 months.
#TACTICS & TOOLS
Tools and techniques I’m using...
Familiar
A Mac app that watches your screen and saves everything as context for your AI agents. Free, open source, offline. Install it and your AI knows what you’ve been working on, which Slack threads matter, what decisions were made in that Google Doc you read an hour ago. AI’s biggest bottleneck is context, and the most valuable context is behavioral: where you spend your attention.
[
Build AI product sense by using AI agents for real work
Introducing Familiar: Never brief your AI again
This week I asked AI…
Read more
3 months ago · 43 likes · 24 comments · Tal Raviv
Proof
An online document editor built for agents and humans to collaborate. Agents suggest edits and leave comments. You accept, reject, or reply. Every character tracks provenance. Paste a doc link into any AI agent and start co-writing. The bet: the best documents will be co-authored with agents, not generated by them.
Fort
The first wearable purpose-built for strength training. Automatic exercise detection, rep counting, muscle-level volume breakdowns, proximity to failure. Every other wearable gives you heart rate and calories. Fort gives you what matters for people who lift: per-muscle insights, form feedback, progressive overload tracking.
#THINGS TO TRY
Experiments worth running...
Clear Your Allostatic Load
Allostatic load is cumulative wear from chronic stress. The reason you feel exhausted even after adequate sleep. The claim: clearing it makes laziness impossible because your body stops running in survival mode. Worth investigating if you’ve been grinding and wondering why rest doesn’t restore you.
Audit Your Meetings as if they were Products
Write one sentence for every recurring meeting: “This meeting is a good use of time if [who] can [what] by [how].” If you can’t write the sentence, cancel it. If the sentence keeps breaking (wrong attendees, wrong cadence), fix the sentence. Most meetings fail because nobody designed them.
#TWEET
Social media that made me stop and think...
AI + Work
Engineers first, but every role is next. Design, product, operations. If sales reps are building production features at Ramp, the “I’m not technical” excuse has about six months left.
The product development loop is collapsing. Idea to spec to prototype to shipped, same person, same day.
Headless services for agents. Email-for-your-agent. Calendar-for-your-agent. The companies agents choose to work with are the next platform winners.
Product + Shipping
The next-gen AI interface might be an infinite canvas. Not chat, not forms. A spatial workspace where agents and humans manipulate the same objects.
Career + Life
The people who thrive try something new before they’re told to. Curiosity and agency, not technical skill.
The most true thing I read this month.