Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
I Found The Layer OpenAI and Stripe Are Fighting Over.
TL;DR
The real AI platform fight is over the “work primitive,” not computer use — Nate argues that OpenAI-style computer control is just a bridge, while durable power comes from defining semantically meaningful actions like “issue a refund,” “reschedule a meeting,” or “authorize a payment.”
He frames agent software as three layers: access, meaning, and authority — browser control and desktop automation solve access, but products win when they also encode what an action means, who can do it, whether it’s reversible, and what happens if it fails.
Coding agents worked first because code already has rich semantics — repos come with tests, type systems, linting, dependencies, and Git history, so tools like Codex and Claude can act, get feedback, and self-correct without a human hovering every 30 seconds.
His practical advice is simple: use the richest interface available before falling back to browser clicks — if there’s an API, MCP, typed object, or connector, use that first, because “guessing is not a strategy for high-consequence work” like contracts, customer emails, deploys, or payments.
Perplexity’s move toward Comet and “personal computer” makes more strategic sense through this lens — Nate says it has to get closer to the browser, files, and workflows where research becomes action, then build a durable work graph across apps rather than staying a search box.
Software companies now face a painful tradeoff between exposing too little and too much semantic access — he contrasts Salesforce leaning into headless, agent-readable APIs and MCPs with SAP resisting agents, arguing that products that hide meaning will be clumsily automated through the UI anyway.
The Breakdown
The shiny demo is not the real story
Nate opens by pushing back on the hype around agents clicking buttons and filling forms. Yes, Codex computer use can move through tabs and do work, but he says that visible activity is distracting people from the deeper platform shift: the future fight is over “who defines what the button means.”
Access gets agents in the room, but meaning tells them what they’re doing
He introduces his core framework: access, meaning, and authority. The calendar example makes it concrete — changing a meeting time isn’t just “click save,” it can notify five people, break a customer commitment, or wreck prep time, and that hidden human context is what he calls a semantic work primitive.
Why browser control is necessary — and still not enough
Nate is clear that computer use matters because most of the world is not API-clean: internal dashboards, Excel workflows, procurement tools, share drives, and government sites still assume a human at the keyboard. But he calls browser and desktop control a “universal adapter” — useful, like screenshots, yet fundamentally shallow, because they expose surfaces rather than underlying structure.
The architectural rule: use the richest semantic interface first
This turns into a practical product rule: use connectors, protocols, typed objects, and permissioned actions whenever possible, and only fall back to browser automation when nothing richer exists. He says this is already how hyperscalers are building — Codex and Claude prefer MCPs and better-structured tools when they can — and he makes a direct plug for users to add plugins and connectors to ChatGPT, Codex, and Claude.
Why coding agents were the first real wedge
He rejects the lazy explanation that coding agents worked just because code is text. The deeper reason is that software development already contains semantic feedback loops — tests, linters, package managers, modules, dependencies, Git history — so agents can inspect state, act, observe failure, and revise without constant supervision.
His startup roadmap: make work legible to agents
This is where Nate gets fired up: if you’re building a startup, don’t stop at shipping an MCP server and calling it a day. Try to break the workflow, find where the agent can reach the levers but still doesn’t understand the task, because that gap between access and meaning is where hard, defensible product value lives.
Perplexity’s browser-and-computer push starts to make sense
He uses Perplexity as the clearest non-hyperscaler example. Moving toward Comet and “personal computer” is strategic because research only becomes durable product power when it sits inside the browser, desktop, files, and app workflows where actual work happens — though he warns that owning the browser is only enough if Perplexity can build a real cross-domain work graph above the apps.
The stack fight: Salesforce opens up, SAP locks down
Nate ends on the broader industry tension: model companies want broad agents, browser companies want orchestration, SaaS companies want to preserve domain authority, and identity layers want control over authorization. His sharpest contrast is Salesforce 360 embracing agents, MCPs, and headless interfaces versus SAP trying to wall agents off; to him, the winning products will be the ones that are semantically legible to both humans and agents.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
