A company running AI agents at scale generates 70 Wikipedias of text per month, roughly 500 years of continuous reading. Every failure is in there. Almost nobody reads it. Not because they don't care. Because the tools don't exist yet. Traditional observability was built for metrics. This is a data engineering problem. Agent traces are megabytes of nested unstructured text. You can't dashboard your way to understanding why an agent went off the rails. @platypii is speaking at @databricks on exactly this. His company @hyperparamapp is worth finding if you operate agents in production. Full blogpost below 👇

and if you're wondering where my hand goes a the start of the video, it's because the office dog came by and could not resist saying hi to Mahina even during an interview

white paper here soffi.ai/how-soffi-works

In this interview I talk about "Beautiful Software", but what is it? "𝗕𝗲𝗮𝘂𝘁𝗶𝗳𝘂𝗹 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲" doesn't rely on flashy marketing, trendy UI tricks, or bloated features to convince the user of its value. The beauty is baked into its foundational architecture, its coherence, and its rigorous alignment with user needs. 𝗜𝘁 𝗶𝘀 𝗯𝗲𝗮𝘂𝘁𝗶𝗳𝘂𝗹 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗶𝘁 𝗶𝘀 𝗳𝘂𝗻𝗱𝗮𝗺𝗲𝗻𝘁𝗮𝗹𝗹𝘆 𝘁𝗿𝘂𝗲 𝘁𝗼 𝗶𝘁𝘀 𝗽𝘂𝗿𝗽𝗼𝘀𝗲. A Motive Force backed company, Soffi just published their architectural white paper. Right now, a designer on your team filed a ticket to change a button label. An engineer will pick it up in three days, configure a local environment, make the change, and send a screenshot for review. Soffi cuts that entire chain. Your whole application spins up live in the browser. A PM points at a component, describes what they want, and an embedded agent makes the change against the real codebase, with tests and type-checking running as it works. A pull request goes into normal review. 𝗦𝗼𝗳𝗳𝗶 𝗿𝗲𝗺𝗼𝘃𝗲𝘀 𝘁𝗵𝗲 𝘁𝗿𝗮𝗻𝘀𝗹𝗮𝘁𝗶𝗼𝗻 𝗹𝗮𝘆𝗲𝗿 𝗯𝗲𝘁𝘄𝗲𝗲𝗻 𝗶𝗻𝘁𝗲𝗻𝘁 𝗮𝗻𝗱 𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻, and with it, entire categories of work that should never have existed in the first place. 𝗬𝗼𝘂𝗿 𝗯𝗲𝘀𝘁 𝗣𝗠𝘀 𝗵𝗮𝘃𝗲 𝗶𝗱𝗲𝗮𝘀 𝗱𝘆𝗶𝗻𝗴 𝗶𝗻 𝗾𝘂𝗲𝘂𝗲𝘀 𝗿𝗶𝗴𝗵𝘁 𝗻𝗼𝘄. Your designers wait days to see a copy change in context. 𝗧𝗵𝗲 𝗰𝗼𝘀𝘁 𝗼𝗳 𝘁𝗿𝘆𝗶𝗻𝗴 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝗶𝘀 𝘁𝗼𝗼 𝗵𝗶𝗴𝗵, 𝘀𝗼 𝗺𝗼𝘀𝘁 𝘁𝗵𝗶𝗻𝗴𝘀 𝗴𝗼 𝘂𝗻𝘁𝗿𝗶𝗲𝗱. ̲𝚂̲𝚘̲𝚏̲𝚏̲𝚒̲ ̲𝚛̲𝚎̲𝚖̲𝚘̲𝚟̲𝚎̲𝚜̲ ̲𝚝̲𝚑̲𝚊̲𝚝̲ ̲𝚌̲𝚘̲𝚜̲𝚝̲. Ask yourself how much of your team's velocity is being eaten by a process that predates the tools you now have. Read how they built it below in the comments 👇