[gavinray]

I recently wound up in conversation with someone working at Amazon. Inevitably, the convo steered into AI/LLM's at some point.

Here is what I learned:

- AWS had an in-house LLM tool that was terrible they tried to use for a while

- A lot of them still use Kiro

- Claude Code is currently the de-facto standard

- They're in the process of getting some custom Codex variant that doesn't phone-home and is audited approved

- There's no mandated organizational standard for what exact tools to some, various teams have different levels of adoption and stacks

- No org-wide/team-wide conventions for Claude Code

- They do have token budgets

- There's an intenral push for something called "Agent Spaces" which was described to me as a sort of Lovable/Bolt-type thing if I understood it right

I can't validate all of this and I might have misremembered, but just in case anyone else finds it interesting.

[pinkmuffinere]

> There's no mandated organizational standard for what exact tools to some, various teams have different levels of adoption and stacks

> No org-wide/team-wide conventions for Claude Code

Just for context, this pattern (different teams using different tools in different ways) is extremely normal within Amazon, and is intentional. These shouldn't necessarily be seen as a failure. Amazon likes to have multiple competing options they use for everything, and they constantly evaluate which option is best performing, like an A/B test. After a couple years they will pare away whatever performs worst, replacing it with a new option. This strategy definitely has it's disadvantages, but it is an intentional chosen pattern throughout the company.

Source: I worked there for 5 years, and painfully/tearfully remember the transitions chime -> slack -> teams and workdocs -> quip -> confluence :')

[nateglims]

You've missed out on migrating to Zoom, Asana, and Sharepoint now.

[antonvs]

How does anyone intentionally decide to migrate to Teams?!

I've only ever seen it in Microsoft shops that used it from the beginning.

[xingped]

Some dipshit with too much power and not enough brain cells nor incentive to care about making good choices. That's also an intentionally chosen pattern throughout the company. :)

[etc-hosts]

Chime was awesome! Also you could shove in Snapchat filters. pointy ears.

[sparrc]

I work at AWS, we got access to Claude Code very recently, most people are using Kiro and that seems to be the defacto standard.

I have not heard of any nor run into any token budgets.

There is generally a lot of team-wide usage of CLAUDE.md/AGENTS.md, team-specific skills, oncall skills, etc.

Not as much org-wide, although my org has an MCP server built for helping with oncall.

[soma8088]

- claude is relatively new, was only whitelisted about 1-2 months ago. kiro is still de-facto, a lot of ai tools use its acp as a back-end.

- there are no token budgets

[noisy_boy]

I am more curious about how much token budget they have. Here I have to beg my boss for more as if he is paying from his pocket or I am using it for my hobby projects (I am not). I guess time to go back to copy/pasting to chat and doing things by hand like a caveman.

[garciasn]

I don't understand this token budget shit. Why would anyone, in their right mind, not be using Claude Max? All of the engineers at our org are using 6.25x and several heavy non-developers are as well. The rest of the company with licenses are using 1.25x.

I have hit my 6.25x limit exactly once in the last quarter.

---

I realize that we will all eventually be forced to pay more for this and I have raised it as a real possibility to the org for budgeting scenario planning; however, for now, why would you pay by token when it's subsidized?!

e: I now understand; you can stop downvoting.

[mrgaro]

In order to get enterprise agreement you need to pay per token for Claude.

[garciasn]

Ah; I didn't realize there was a 150 person cap to Team and I suppose paying out the ear is worth it for compliance audit. Makes sense in that regard.

[disgruntledphd2]

> I realize that we will all eventually be forced to pay more for this and I have raised it as a real possibility to the org for budgeting scenario planning; however, for now, why would you pay by token when it's subsidized?!

Anthropic (and maybe OpenAI?) have gated all the important enterprise features behind API plus pricing in the last quarter or two.

[mrguyorama]

For the code I generate and the limited way I am using it, Claude Sonnet is reliable and good.

I hope that I can someday run something very much like it locally.

The moment that happens, the AI industry is essentially useless to me. I don't need some ultra expensive "Totally better" model that does the exact same thing.

[dastbe]

because they’re not; anthropic is pushing enterprises to switch to API/token pricing

[8note]

assuming it hasnt changed, agent spaces is just a developer desktop running on actual aws rather than on internal hosts, and they figured out how to put vscode on it in a security approved way, including your employee credentials.

one of the things that allows is for adding mcps, skills, and various harnesses that are preconfigured to work out of the box.

i doubt its gotten out of the employee needing to sign in every couple hours

[tyingq]

Would be really curious what the internal market share for Kiro is. Not a really good look for it if it's just smattering use here and there.

[siva7]

It's the worst harness i ever used. And they are selling it like crazy to all their enterprise customers who don't know any better. A true money printer at this point.

[nh23423fefe]

i like it a lot. why dont you?

[Ancalagon]

> There's an intenral push for something called "Agent Spaces" which was described to me as a sort of Lovable/Bolt-type thing if I understood it right

Everyone and their mother making vibe-app platforms for their company API's now

[add-sub-mul-div]

From the outside it always seems laughably circuitious compared to just learning skills ourselves.

Even if it did let me fill out TPS reports 20% faster, who even cares compared to all of this chaos?

[dfee]

the problem is scale. there's a tension between an individual developing technical skills (transfer cost is high, slow, expensive) and developing agent skills (transfer cost is low, instant, free).

so, just like a manager manages employees, or you consult a contractor, agents are a way of getting leverage over a system.

that said, if you want to learn to play saxophone, you're free to do so. just note your personal endeavors may begin to look more like hobbies than marketable skills.

[rustystump]

When the cost for this leverage is more than an employee the math stops mathing.

Additionally, for tech work. There is a tension about doing work and not knowing that output is correct or not. I have seen ai spit out thousands of lines of opencv code for a simple color lut. The person doing this had no idea what was going on. If they continued, the token cost and time waiting for agents spinning only goes up.

Yes, agents get smarter and cheaper but the above example replays over and over again even on crud apps. You still need to dev the skills and transfer costs for it to be effective.

[wahnfrieden]

Shareholders want the company to figure out how to pay fewer employees, and pay them less by down-skilling them