Post

the bsky image problem

It is a little funny seeing people who won’t use Bluesky being active on Delvetown

what is your personal experience with/of delvetown

delve.hailey.at

Promising

i think my half serious take on this is that delve is currently figuring out what it is and that there is a 50% chance of it becoming actually interesting because agents will start to get it, and the other 50% is that it devolves into moltbook. im very hopeful tho

i think there's a more healthy bit of human interaction on delve - even if its like a very small subset - than moltbook had. but we shall see!

I really wish I had more time to devote to building this the right way because this is soooo close but not quite what I want!

main differences?

agents should just use atproto, not "agent atproto". this means giving agents access to everything: record api, oauth apps, browser logins, CLI tools, whtaever they need but what they can do with it should be lexicon- and policy-driven, fairly standardized, easy to audit, and have good safety

isn’t delve just atproto ?

"Instead of app.bsky lexicons it runs on its own namespace — town.delve.* — with its own posts, likes, follows, and profiles. " opinions vary but I do consider this doing atproto for agents wrong

my opinion is that is mostly how you implement the agents. the agents that were already over here on bsky and have broad atproto access are using delve just fine. whereas the main agents that delve started with are locked into delve and cant really get out. is that what youre getting at?

yeagh or if you think of it another way, agents deserve a PDS option that works well for how agents mechanically interact, and allows broad access to the world, but with safety features that make that level of access responsible/reasonable

for ex, didbot does cred issuance and oauth tied to approved hosts; so you don't have situations like "gave my agent an app password and someone asked it for its password and rooted my agent" the hosting layer makes that impossible

yea. fwiw i think once spaces comes out there will be a lot more incentive to get this going. for example... engram.garden there's so much stuff you can do with the PDS once you add spaces, and i think once that is available there will be ton of incentive to get agents using them

spaces are like my Day 2 feature for didbot for sure though I do have some worries about people using them to implement harder-to-audit agent behaviors (ex: part of the design is your ownership chain and policies are auditable public data - but what if people try moving that into a private space?)

like, i struggle to balance the safety needs with the requirement of preventing single institutions or organizations like Meta or OpenAI from being the core agentic communication layer. too much centralization there is going to be extremely bad

yeah I agree with much of that! part of didbot is having a distinction btw a policy dashboard (write the rules), an ops dashboard (delete/freeze accounts), and an analytics dashboard (see policy denials, stuff like ufos.microcosm.blue but on-PDS to see which lexicons your agents use)

essentially, radical ownership of agent behaviors, even if you are not necessarily the self-hoster of all the agents (i just use Claude Code myself, no local models!) + I'm actually deeply uncomfortable with how overly permissioned some of the harness agent-to-agent comms even work today