CLI, Skills, Vertical Models: The New Software

In the era of Agent Experience (AX), AI agents—not human users—are emerging as the primary users of software. Agents don't tap dashboards; they work through APIs, scripts, and structured commands, so products need to be rebuilt around stable interfaces, CLI, MCP servers, skills, and, where needed, faster vertical models, rather than GUIs.

In January 2025, there was talk of the "death" of software, but the conclusion by April 2026 is a bit different. SaaS hasn't disappeared—rather, the human-centered usage flow has collapsed, and designing products for the agentic user has become the new standard. In enterprise environments, machine identities now outnumber human users by 45:1, with some reaching as high as 100:1, and Neon reported that 80% of DB creations come from AI agents. On GitHub, commits written by Claude Code are said to account for over 5% of all commits, with AI-assisted commits at around 40%.

The core idea isn't "bolting a chatbot onto a dashboard." Agents don't want a UI that a person clicks through directly—they want readable inputs and outputs, retryable commands, tool chains, and structured responses. That's why the first entry point for a product becomes the terminal, not a screen.

Three patterns are building the new stack.

• Skill files: Encode domain experts' judgment into markdown so agents use the product in the right order and within the right constraints.
• CLI / MCP servers: An interface layer that lets agents handle installation, configuration, and operations.
• Vertical models: Models trained on specific domain data, making them faster, cheaper, and—for certain tasks—better than general-purpose LLMs.

Figma released Skills alongside its MCP server, providing design system rules, component naming, and token structures as files. PostHog is rebuilding its agent architecture from scratch, moving practical knowledge into skills—for example, guiding agents to default to $pageview rather than signed_in as the activation event. Without this kind of documentation, agents can produce analyses that look plausible but are wrong.

CLI is not an old-school developer tool but an agent-friendly interface. 37signals rebuilt Basecamp into an agent-accessible product complete with an API, a new CLI, structured JSON output, and shell completion, and Google Gemini CLI extensions reportedly secured 1M+ developers within 3 months. Vercel AI SDK has also surpassed 20 million monthly downloads, establishing itself as central to agentic pipelines.

The third pillar is vertical models. Dedicated models tuned for specific reasoning, like Intercom's fin-cx-retrieval for customer support, outperform general-purpose models on specific problems. Cursor Composer 2, built on Kimi K2.5 with additional pre-training and RL, scored 61.7% on Terminal-Bench 2.0, surpassing Claude Opus 4.6's 58.0%, while being offered at around $0.50 per million input tokens.

However, vertical models don't always create a moat. Harvey achieved 97% lawyer preference with a legal model trained together with OpenAI, reaching $190M ARR and an $11B valuation, but after frontier reasoning models from Google, xAI, OpenAI, and Anthropic outpaced it on its own benchmarks, it retired that model and began routing between Claude, Gemini, and GPT through a Model Selector. When baselines improve rapidly, an advantage baked into weights can vanish easily.

In the end, the strategy that survives is combining all three.

• A weights-based data advantage for the most expensive queries
• Skills and workflow expertise for task procedures
• CLI/MCP-based composability for execution and automation

From a performance standpoint, the answer is also clear. Most production-environment tasks don't require frontier reasoning. Work like contract extraction, data validation, format conversion, and status checks can be handled by code or small models, and as FrugalGPT research showed, cascade routing can produce up to 98% cost savings. In agentic workflows where 5 to 15 calls chain together, whether every step runs through a large model or not can be the difference between 30 seconds and 2 seconds of latency.

Software's competitiveness is now determined less by button colors or screen layout, and more by ease of authentication, security, speed, and cost. Going forward, SaaS will favor re-encoding data, workflows, and domain knowledge into forms that agents can read and execute, rather than human-facing GUIs.