Three Layers Define Agent Power: Access, Meaning, Authority

Agents interact via access (computer use like browsers/desktops to click buttons), but this is merely a 'universal adapter' for legacy human-built software—shallow and guess-prone for high-stakes tasks. True power lies in meaning: semantic work primitives that encode task context, like a calendar invite's ripple effects (notifications, conflicts, commitments) beyond 'click save,' or a 'buy' button's implications (fraud, fulfillment, disputes). Authority adds governance: permissions, reversibility, approvals (e.g., read vs. write, draft vs. send, sandbox vs. production). Without meaning, agents guess wrongly on refunds, deletions, or emails; humans intuitively grasp this, but software hides it behind forms. Control these layers to reduce supervision—trusted actions aren't binary but nuanced by semantics.

Practical fix: Follow the hierarchy of richest interfaces—APIs/connectors first, then protocols/typed objects, fallback to browser/desktop. Plug in MCPs, plugins to ChatGPT/Claude/Codex for better results; this exposes structure over screenshots.

Coding Agents Succeed First Due to Rich Semantics

Coding agents arrived early not just because code is text, but because codebases offer dense semantics: modules, dependencies, tests, linters, git history provide feedback loops (run test, see error, revise). Tests aren't verification—they're meaning artifacts signaling the 'world' (e.g., staging vs. production). This lets agents self-correct without constant human input, unlike knowledge work (strategy docs lack tests; calendars hide politics/relationships; sales/procurement rely on unwritten history). Coding is a 'wedge' for agent-native software: expose primitives like refunds, reschedules, meeting briefs directly, making non-coding work legible.

Outcome: Agent-native systems minimize human coordination; startups should map semantic gaps in MCPs/APIs to build moats—solve where prompts fail due to missing task understanding, avoiding errors like bad tones, wrong refunds, or inconvenient invites.

Platform Strategies Reveal the Moat Fight

Hyperscalers (Claude/Codex) start from models/code semantics, composing tools effectively but struggling with real-world purpose (e.g., calendar conflicts). Non-hyperscalers like Perplexity work backward: from search to browser (tabs assemble cross-app context: email/docs/SaaS) to computer/files for workflows (e.g., finance in Personal Computer), building durable 'work graphs' above apps with permissions/validation. Trap: Stay operator (just interfaces) vs. assembler of meaning.

Enterprise signals: Salesforce goes headless (exposes semantics), SAP blocks agents (guards meaning). Leaders err asking 'can it act?'—ask 'does the product know what the action means?' Demos distract; build for primitives where model + harness + legible work = autonomy. Commerce hint: Agentic transactions need semantic layers (discovery/checkout/infra).