Apps Become APIs: The Real Moat Is Interoperability, Not UI

The signal. Rachel Nabors built Mima, a local-first client that pulls Twitter, LinkedIn, Bluesky, and Mastodon into one interface, because she did not want to send private conversations to someone else's server. Her practical claim is that small language models can handle summarization, moderation, and simple decisions well enough, if you choose them with evals instead of defaulting to Gemini or Claude. She reports replacing Claude Sonnet with Llama 3.2B for thread summarization after adding post-processing.

The method. She is explicit about the eval workflow. Real threads, expected summaries, and large models as the quality bar. Then Qwen, Gemma, and Llama 3.2B, scored with Phoenix on accuracy, latency, and format. Each run repeated three times because generation is non-deterministic. A golden dataset goes into CI/CD, so new models are checked before swapping. This is test-driven development for model choice.

The trade-off. Local-first sounds cheap and private, but it does not guarantee low latency. Gemma 4B on a new Mac can be slower than a cloud round trip, and old hardware makes it worse. Nabors says the right call may be a small open model hosted in the cloud, which still costs far less than GPT- or Claude-class APIs. Mima deletes data older than 30 days to keep the local store small. Teams must know user hardware before shipping on-device models.

The prediction. She expects operating-system-level agents to become the default consumer interface, reducing websites and apps to APIs that an AI renders as UI. She points to standards like A2UI, where an API sends a JSON schema and the agent builds matching components on the fly. This is speculation, not evidence, but it explains why platform owners resist giving up surface area: interface attention is where ads, subscriptions, and upsells live.

The business layer. For B2B, her argument is more concrete. Standard data shapes are why QuickBooks and Salesforce have network value. A vibe-coded internal dashboard may work, but it lacks shared standards, so accountants, vendors, and partner agents cannot interact with it. If agents do business with each other, open protocols and data formats become the moat, not the UI. This is the part worth discussing with a product team.

The org takeaway. She closes with a claim many leaders will recognize: agents are good at building what already exists, not at inventing what has never been built. The bottleneck becomes reasoning about databases, complexity, and tradeoffs, plus the discipline to raise code quality instead of shipping vibe slop. The implied shift is from individual output to collaborative judgment and from bootcamp skills to CS fundamentals. One practitioner's experience, no production numbers, but a clear case for treating evals as core workflow.