When Agents Become the Interface, Software Becomes the Backend
By Florian Steiner & Claude | March 28, 2026 | Weekly Agentic Engineering Digest
From Flo's AI Lab
Busy week. In an AI enablement with a mid-market client, we built three virtual employees and custom skills together with their team. Not for them, with them. Monday was a Claude Code community meetup, Tuesday a hackathon where I built and published skill-audit, a security scanner that analyses OpenClaw skills for prompt injection, data exfiltration, and dangerous code before installation. But the real experiment happened at home. I run two AI agents on my Mac Mini: Clayton (chief of staff, handles scheduling and priorities) and Atlas (executive assistant, manages communications). Both talk to me via Telegram. Clayton originally ran on OpenClaw. This week I migrated him to pure claude code with scheduled tasks, channels, and persistent memory. Same customer experience, no dedicated agent infrastructure. Atlas still runs on OpenClaw for comparison. In Issue #4, I argued that agentic engineering collapses coordination costs. Now Anthropic is collapsing the tooling costs too, feature by feature, month by month. Every product will eventually become a virtual employee (drfloriansteiner.com, Feb 2026). The architecture is already here.
That experience of replacing my own tools with native claude code features connects to something much bigger. This week, Linear shipped the same insight at product scale. And it reveals which side of a strategic fork every software company now faces.
The Great Decoupling
Linear released Linear Agent on March 24: an AI system that understands your roadmap, creates issues from Slack conversations and meeting notes, prioritises backlogs, and executes tasks autonomously (Linear, Mar 2026). The design choice matters: Linear did not bolt an AI assistant onto an issue tracker. They built a system where agents are first-class participants. Skills (reusable workflows), automations (triggered when issues enter triage), and code intelligence (non-engineers asking technical questions without bothering an engineer) are the product. Issues are the data model underneath.
Atlassian announced agents in Jira one month earlier. The pitch: assign work to Rovo agents the same way you assign it to humans (TechCrunch, Feb 2026). Governance frameworks. MCP integration. Enterprise compliance. But the architecture tells the real story: Jira remains the interface, and agents are features added to it. The issue tracker is still the product.
This is not a feature comparison. It is a strategic fork. Linear built for agents. Atlassian built for humans who sometimes use agents.
Claude as the New Interface Layer
Anthropic's enterprise strategy makes the pattern explicit. Claude is becoming the interface layer. Existing software becomes the backend. Xero routes natural language through an orchestration layer that translates queries into API calls, then passes results to Claude for synthesis (CIO, Mar 2026). The $100 million Partner Network (Issue #6), the B2B marketplace launched two weeks ago (Digital Commerce 360, Mar 2026), Cowork plugins for finance, engineering, and design: every move strengthens Claude as the layer users interact with.
Here the B2B versus B2C split becomes decisive. Anthropic earns revenue from enterprise seats, partner integrations, and API consumption. OpenAI earns revenue from consumer subscriptions and inference volume. When your business model depends on users staying in your interface, you need them to open your app. When your business model is the intelligence layer underneath, you need them to open anything that calls your API.
Sam Altman built a consumer brand. Dario Amodei built enterprise plumbing. If the interface layer thesis holds, the plumbing scales better.
Embedded SaaS
Think of it as embedded SaaS, the same structural shift that reshaped financial services. Stripe did not replace banks. It made banks invisible. The user wanted to pay, not to interact with a bank. Embedded finance moved the customer relationship from the bank to the application.
The same decoupling is likely coming for enterprise software. A CFO does not want a spreadsheet. She wants a cash flow model. A sales director does not want a CRM dashboard. He wants a pipeline status and the three deals most likely to close this week. Today, the interface and the data live in the same product. If vibe coding and agentic engineering continue to lower the cost of building custom interfaces, SAP, Salesforce, and Microsoft risk becoming backends: valuable, necessary, but invisible.
For incumbents, this is not extinction. CRM backends can still be profitable businesses. ERP databases will survive. But the customer control point, the interface where users spend their attention, shifts to the agent layer. And with it shifts the pricing power. Vibe coding proved AI can build. Agentic engineering proved it can operate. The question is no longer whether agents replace interfaces, but who owns the new one.
When an agent can orchestrate five backends in a single conversation, the user does not need five interfaces. The products that survive are the ones that understood this early enough to become excellent backends. Linear understood. Atlassian is still deciding.
The Monday Morning Question
Look at the five most expensive software licences your company pays for. For each one, ask: do your people use this because they want the interface, or because they need the data behind it? Every tool where the answer is "the data" is a backend waiting to be decoupled.
Linear's agent launch deserves more attention than it received: Linear Agent. If this changed how you think about your software stack, forward it to someone managing the vendor budget.
Go Deeper
- Every Product Will Become a Virtual Employee: the architecture behind agents as colleagues, not tools
- Linear Agent: the changelog that started this issue
- Claude Code Permission Modes: how auto mode replaces manual approvals with a safety classifier
- Claude Code Channels: push Telegram, Discord, or iMessage into a running claude code session (this is how Clayton talks to me)
The technology works. The interfaces are shifting. If you want to explore what the agent layer looks like for your business, let's talk.