Almost every founder I know has one line from a customer that's better than anything they'd have written themselves. Usually it's the oddly specific one. The bit you'd never have thought to claim about your own product.
What I find interesting is how those actually arrive, because it's almost never the polite follow-up email asking for a testimonial. Mine turn up by accident. Someone says it in passing at the end of a call, or drops it into a support thread, and if you don't catch it right there it's gone for good.
Spent the last stretch wiring MCP servers into an agent product instead of hand-building every integration. The pitch is great in theory: paste a URL, click connect, tools show up as a checklist.
In practice it's been a mixed bag. Some servers are exactly as plug-and-play as advertised. Others expose tools that don't match how the underlying API actually behaves, so the agent confidently calls something that returns garbage, and you don't find out until you're debugging why a real user got a wrong answer.
Ended up hand-building the connectors that really matter (the ones where getting the semantics wrong costs someone money) and using MCP purely for the long tail where "mostly works" is good enough.
Anyone else gone deep on MCP servers? Curious if others found the same gap between the demo and production, or if I just hit unlucky servers.
I've gone back and forth between these two more times than I should probably admit.
Coda always pulls me in because I can make tables, buttons, little automations, etc. It's fun until I've somehow turned a simple notes system into a project of its own.
Then I go back to Notion because I just want to write things down and find them later.
And then, inevitably, I miss something Coda did better.
We read both vendors' documentation and the recent work on looped transformers with three questions: what does a model's reasoning cost, what can you steer, and what can anyone read afterwards. The answers sort into three places.
On the page, as visible text. You pay the output rate, you steer it with the prompt, and you can read it, though studies show these visible traces are often unreliable.
Off the page, as hidden tokens. The example in OpenAI's own guide shows 1,186 output tokens charged, 1,024 of them reasoning; you pay for every one and the text stays with the vendor. Anthropic's docs show the same: you're billed for full thinking tokens whether shown or not. The rate is the same, the steering is an effort knob in place of a prompt, and what you can read is at most a summary. The exact count is in the usage object, and on the newer models earlier turns' hidden reasoning stays in context and is billed again as input.
Inside the network, as extra passes through the same layers, which is what recurrent depth means. Compute rises per token, the token count does not move, and there is no text for anyone to read, the vendor included.
