pull down to refresh

20 sats \ 2 replies \ @UncleJim21 10 Jun \ on: (L)402 is a literal failure code, btw. bitcoin lightning

Hey Justin - I've been working in the agent space for a bit. I actually just published this if you're curious: https://www.pullthatupjamie.ai/for-agents.

I share your skepticism that L402 is "the thing". I adopted it out of desire to help bitcoin and to help us win the "arms race" of best rails for agents. A few shortcomings I've noticed as well. A lot of L402 assumes statelessness and "onesy twosy" atomic actions. That precipitates two problems:

  1. The overhead of each request triples. I have to:
    a. Hit the endpoint knowing im getting a 402
    b. Pay out of band
    c. Hit the endpoint with my macaroon
  2. It assumes that endpoints are unrelated. In my case, I have about 4 or 5 endpoints that are related (search different levels of podcast corpus hierarchy, ask a question in natural language etc). That's clearly not the case and it would be silly for me to NOT keep some persistence for user convenience as their agent digs around in our 125k+ hour podcast library.

I think you can get around that with caveats. I personally got around it with an optional credit system where they can prepay for X sats worth of usage and reuse the macaroon. Some people bitched about that because they felt like I was ripping them off by recommending the "batched" credit approach.

You are right about the identity point. But the problem/pain for users is not really identity, it's the inefficiencies we incur by not having optional persistence (identity is one solution).

write
compose
47 sats \ 0 replies \ @UncleJim21 10 Jun

@justin_shocknet love the discourse - we should chat live sometime. Hmu on nostr anytime or i reached out to you in a slack group i think we're both in https://primal.net/p/nprofile1qqsy2d24rfqzwxc9n2ujku084dlgwqqxrgkervxjpucna7p0q5htppgeluehw

reply
3 sats \ 0 replies \ @justin_shocknet OP 10 Jun

Yea I think it ultimately comes down to the pattern of API keys isn't going away for the reasons you described, a macaroon is still like an API key but with extra steps in most cases. Sounds like your persistent of them follows that pattern, seems inescapable to do so.

If we're going that route, may as well use an nsec for the other benefits. I think it'd be cool to wrap your endpoints in a CLINK offer, would be an interesting experiment in UX to use them directly from Nostr clients... will hit you up on slack.

reply