AI agents can now open browsers, move through tabs, fill forms, change calendar events and manipulate files. But Nate B. Jones argues that this visible capability is not the real platform shift. The future is not an AI that gets better at clicking buttons; it is software that can explain what those buttons actually mean.
The key concept is the “work primitive”: a semantically meaningful unit of work such as issuing a refund, rescheduling a meeting, authorizing a payment or preparing a meeting brief. A button is not just an interface element. It can involve money, consent, customer data, obligations, permissions and risk.
Jones frames the problem in three layers: access, meaning and authority. Access gets the agent into the workspace. Meaning tells it what it is touching. Authority defines what it is allowed to do, what can go wrong and how the outcome should be checked.
Key points
- Computer use is necessary, but it is a shallow interface when used alone.
- Agents should use the richest semantic interface available: connectors, protocols, APIs, MCP servers or typed objects before falling back to browser or desktop control.
- Platform power will accrue to companies that define and expose meaningful units of work.
- Coding agents arrived early because software development already has unusually dense semantics: tests, types, dependencies, linters, package managers and Git history.
- Trust is not a simple read/write switch: agents may stage but not deploy, recommend but not approve, write in one space but not another.
- Future software must remain simple for humans while making objects, operations, permissions and consequences legible to agents.
Why it matters
- Moving a calendar invite can affect commitments, relationships, preparation time and priorities; the agent needs more than field access.
- In refunds, contracts, customer emails, payments or deployments, guessing is not an acceptable strategy.
- SaaS companies face a strategic tension: expose too little semantic access and generic agents will operate clumsily; expose too much and the product may become backend infrastructure for someone else’s interface.
- The coming platform fight spans models, browsers, SaaS tools, identity systems and permission layers.
- Salesforce and SAP are presented as contrasting examples: one leaning into agent access, the other appearing more defensive.
Signals to watch
- AI browsers that can build context across calendars, email, documents, dashboards and SaaS applications.
- Whether Perplexity, Comet and similar products can turn search into structured, permissioned and reviewable action.
- Broader adoption of MCP, connectors and richer APIs in everyday work software.
- Products that expose what an object is, who owns it, what action is proposed, whether it is reversible and how success is verified.
- Hyperscaler moves around Codex, Claude and agents that operate close to code, files and compute primitives.
- A shift from agents that merely operate interfaces to systems that understand the work behind the interface.
Source
- Chaîne: AI News & Strategy Daily | Nate B Jones
- Vidéo source: https://www.youtube.com/watch?v=b1fxYGPbHeo
No comments yet