Stripe is the company I usually hold up as the exemplar of polished internal tools. Hopefully this is taken the right way and I don't want to be negative towards the teams working on this, but I see a distinct lack of polish in these tools and presentations. Some examples:
- Unnecessary AI copy throughout the interfaces like "Browse, discover, and manage skils for your agents", "No favorites yet — hover a card and click the star to pin it here", "One execution environment, shared across agents". These instantly read as AI copy and decrease my enthusiasm.
- Inconsistent, AI-sloppy look-and-feel with different typefaces spattered across the interface
- The session metrics slide looks busy and AI-generated. It repeats 360,014 sessions in one of the cells at the top, but also has a "360K sessions" in the heading
I don't know if I'm the only one that notices this stuff, or whether others see it too.
They have published a bunch of stuff about their internal tools in the past. Look up https://stripe.com/blog/stripe-home and compare it to this, as an example.
It's always possible that in reality they were always a bit less polished behind the scenes though.
I want to read this, but the ever-changing linear gradient of the background is too visually distracting. I tried to get around it by highlighting text I want to read, but since the background is changing underneath it, so is the highlight. My eyes hurt.
Very cool demonstration of managed agents built for the needs of their own business.
I think this is where a lot of companies are going to go: on-prem platforms that give various teams access to agents that are as powerful as coding agents, but much more managed and governed.
I read a buzz word "Knowledge AI Platform" but I did not see any specific feature helpful for knowledge management like verification or transparency. It is more like any generic Agent builder. Maybe it meant to justify building something internally.
I think that’s because none of them go past “I’ve set up agents to be orchestrated this way” and that’s about as impressive as “look at my cloudformation template”.
It's very close to the direction we're taking for windmill.dev, we call it "operator builders" rather than focus on "knowledge graph" which the article is very light on details of.
What I'm mostly reading is a developer platform and runtime where users can build agents that can run tools and for that you need a secure code runtime, ACL/permissions, easy way to build apps or what cloudflare OS call gadgets. We're betting on this too at https://github.com/windmill-labs/windmill, very curious to see if that's the future for most enterprise and if a model where everyone vibe-code/fork cloudflare OS to their enterprise need is the future, or a more exhaustive/enterprise platform like ours does.
Ahhh I built an internal set of agents to run our company (finances, all info in the karpathy-style LLMWiki and a database of clients, contracts, billing, time tracking, all managed by MCPs, etc) and it's also called Kai (company name is Kaizen)
TIL people still fall for eh I mean use Langchain. Sorry, low value comment; I don’t know how to do that differently; it is such bad garbage since day one and strangely it did not improve. Sorry anyway for the comment, at least it was not LLM generated?
It (and most other systems) abstract the wrong concepts when you want to create an agent. They promote what was likely never a great strategy but even parsimoniously what are strategies that were good ideas months or a year back.
For example how langchain or whatever else handles subagents and deep research is laughable even today.
So whats the recommendation? Just use pydantic and code everything up yourself. Maybe strands I dont know.
Which specific strategies do you think are outdated, and what would you recommend instead? This sounds more like a criticism of older versions of LangChain than its current APIs. The post also uses deepagents, a separate package in the same ecosystem. Is there something in the current implementation of either that you’re referring to?
The fact that Stripe are touting a production product that (at face value) benefits their business seems to suggest that maybe crap is subjective and as ever, being overly opinionated in an emerging space might actually be a blocking mindset rather than a positive one.
It's a lot of sugar and high level abstractions on top of existing things, I find using the existing things not that complicated or difficult and I struggle to see the value on using the framework as it doesn't sit well with my way of abstracting the stack.
Size and scale, I would think. An out of the box solution probably doesn't quite have the same capabilities as something they can (and now have to) manage in it's entirety.
Stripe probably WANTS to be opinionated about how their company works with the tools.
Somewhat interesting how the design of Stripe's developer-focused landing pages used to be the "cream of the crop," so to speak, but now they just look like Claude run amok. I wonder if the web design "skills" for Claude, and other LLMs, were overly trained on Stripe.
1. The whole interface feels way too vibe coded with tons of unneeded stuff. Why does it have a console and snake game built in? Why does it have annoying sounds? Why is it full of AI slop writing? The latter is especially confusing because the first person listed under authors is a "Technical Writer". I guess the interface is the general stripe.dev page not exactly related to Kai but the post definitely is pure AI slop writing.
2. It seems to claim things that might not be substantiated like the following: "When Account Executives use Kai, they produce 2x the sales activity, create 17% more opportunities, generate 26% more revenue opportunities, and close 39% more deals when compared to the same sellers in weeks they don't use it." Correlation is not causation. More engaged, better sales people might just also use more tools like this one, it doesn't mean all the increase in sales is because of the tool.
3. the post talks a lot about how great this tool is but it describes nearly nothing of value to the outsider who can't access it. What were the valuable lessons learned? What's neat about it? There is not much meat imho.
It doesn't live up to my usual expectations from Stripe.
- Unnecessary AI copy throughout the interfaces like "Browse, discover, and manage skils for your agents", "No favorites yet — hover a card and click the star to pin it here", "One execution environment, shared across agents". These instantly read as AI copy and decrease my enthusiasm.
- Inconsistent, AI-sloppy look-and-feel with different typefaces spattered across the interface
- The session metrics slide looks busy and AI-generated. It repeats 360,014 sessions in one of the cells at the top, but also has a "360K sessions" in the heading
I don't know if I'm the only one that notices this stuff, or whether others see it too.
unless you work there how would you know this?
It's always possible that in reality they were always a bit less polished behind the scenes though.
I think this is where a lot of companies are going to go: on-prem platforms that give various teams access to agents that are as powerful as coding agents, but much more managed and governed.
I'm betting on this with my open source project, Lightspeed: https://github.com/smartcomputer-ai/lightspeed
I think that’s because none of them go past “I’ve set up agents to be orchestrated this way” and that’s about as impressive as “look at my cloudformation template”.
What I'm mostly reading is a developer platform and runtime where users can build agents that can run tools and for that you need a secure code runtime, ACL/permissions, easy way to build apps or what cloudflare OS call gadgets. We're betting on this too at https://github.com/windmill-labs/windmill, very curious to see if that's the future for most enterprise and if a model where everyone vibe-code/fork cloudflare OS to their enterprise need is the future, or a more exhaustive/enterprise platform like ours does.
By force or by choice?
Happy to be in such good company!
The agent has switched a few times, initially NanoClaw, then Hermes, and now Vercel's Eve (maximum customizability).
NextJS/Shadcn web app, postgres db, eve agent layer, MCP tools (over 100 so every single thing can be done by an agent), SwiftUI mobile app, etc.
Integrations with gmail, gcal, gdrive, quickbooks, using mdx for the wiki displays (so we have rich diagrams, 3d models, etc).
We write about it here: https://kznconsulting.com/work/how-we-run-kaizen
For example how langchain or whatever else handles subagents and deep research is laughable even today.
So whats the recommendation? Just use pydantic and code everything up yourself. Maybe strands I dont know.
I use Apache Burr but it's not as good as LangChain. I just don't want to use software whose website has a pricing page if it's not a SAAS.
seems every company that has spare engineering resource all builds such thing internally
> No existing tool could handle the data security requirements and specific workflows Stripe needed
Stripe probably WANTS to be opinionated about how their company works with the tools.
Spotify built some nonsense like this too and made ridiculous claims that it saves 90 percent of tokens. I guess anything goes these days.