- Published on
Will Everyone Work Inside the Agent?
- Authors
- Name
- Adrian Gan
- @AdrianGanJY
The Question
Here's a thought experiment: if the formula for an AI agent is IDE + Context Pack + CLI Library, and the agent can do everything a custom-built web app can do — will people eventually spend more time talking to their agent than clicking through software?
I think yes. And I think the implications are bigger than most people realize.
What's Already Happening
The shift from "use software" to "talk to AI" isn't theoretical. It's observable:
- ChatGPT has 200M+ weekly users. People are comfortable talking to AI.
- GitHub Copilot has 15M+ developers. Programmers already shifted to AI-assisted workflows.
- I already work 80%+ inside an agentic IDE. It's my primary workspace. Not a browser. Not an app. The IDE.
The direction is clear. The question is speed, not destination.
The Convenience Threshold
People don't adopt tools because they're faster. They adopt tools because they're more convenient (方便). Speed is a bonus. Convenience is the requirement.
So the question isn't "is CLI faster than GUI?" The question is: "is asking the agent more convenient than clicking through a portal?"
The answer depends on the task:
| Task Type | Agent vs GUI | Why |
|---|---|---|
| Single-action (approve invoice, mark complete) | GUI wins | One click beats a sentence |
| Multi-system (check warranty + draft reply + log ticket) | Agent wins | Agent chains 3 CLIs in one request |
| Variable/judgment (draft customer proposal, compare options) | Agent wins | Agent handles ambiguity; GUI needs exact inputs |
| Discovery (browse, explore, scan) | GUI wins | Visual scanning is faster than asking questions |
| Repetitive template (same form, different data) | Tie, then agent | Agent can batch; GUI needs manual repetition |
The crossover: agents win when the task spans multiple systems or requires judgment. GUI wins when the task is a single click or visual scan.
As CLIs cover more workflows and context packs get richer, the "agent wins" zone expands. The "GUI wins" zone shrinks.
The MIBOS Paradox
Here's the uncomfortable thought I've been sitting with:
If I build a full MIBOS frontend — beautiful dashboards, polished UX, the whole thing — and then 18 months from now, agents are good enough that my team prefers asking over clicking... I've built a GUI that becomes a relic. The investment becomes a sunk cost.
But if I DON'T build the GUI, and agents aren't good enough yet for non-tech users... they have no tool at all.
Resolution: Build the CLI now (it's needed regardless — the agent uses it). Let the IDE market catch up. Within 12-18 months, there'll be a "Claude for non-developers" — a simple interface where anyone can talk to an agent without touching a terminal. My CLIs and context packs are ready to plug in. I didn't waste months building GUIs that became relics.
Build What's Yours, Buy What's Generic
This is actually a build-vs-buy decision (自建 vs 外购):
| Layer | Build or Buy? | Why |
|---|---|---|
| CLIs | Build | Your business logic, your unique workflows. Nobody else builds a Mipos HR CLI |
| Context Packs | Build | Your company knowledge, your SOPs, your principles. This is your moat |
| APIs | Build | Your data, your rules, your domain |
| IDE / Interface | Buy / Wait | Generic problem being solved by Claude, Google, and others. Let the market compete |
Build what's unique to you. Let the market build what's generic. This is the same principle that made SaaS successful — don't build your own email server, use Gmail. Don't build your own agent interface, use whatever the market produces.
The Implication for Today
- Stop designing GUIs for internal tools. Design CLI specs and context packs instead.
- Train people to work with agents, not with software. The skill isn't "how to use the HR portal." It's "how to ask the right questions."
- Your CLIs are future-proof. Whatever IDE wins the market, your CLIs plug right in. Context packs are markdown files — they work everywhere.
- The only non-negotiable build is the API + CLI. Everything else is optional, temporary, or someone else's product.
The future of work might not be better software. It might be no software — just agents, context, and CLIs.
Adrian Gan is the CEO of Mipos Sdn Bhd. He's currently betting that the best interface for business software is a conversation, not a dashboard.