dhorthy

i am very long on TS as the future of agent applications. nice work team

interesting take - thanks for reading our story- here's to the future of human-aided-bash-novices not munging our PATHs up

I think these are fair comparisons to make. I think the value is less in the slack api integration - anyone can plug in a slack client in afternoon.

When you get to timeouts, escalations, and routing 100+ conversations between 4+ users across multiple slack instances, that's when things got hectic for us.

is that because of more deterministic AI flows like llm-as-judge, rag reranker, post-eval, etc?

do you think something like langgraph state is sufficient?

i like this idea - runtime inference based on past responses that gets smarter dynamically is a really interesting space

we are certainly out here

lol this is why i come to this site

you get it

lol i like LLord I might steal that

awesome - glad to hear it resonates

BRB rebranding to meatgap.dev

yeah I think that's right - we put humanlayer in between the non-deterministic (LLM Decision) and the deterministic code (tool execution logic)

hmm, like, i love simplicity, and I'm open to other approaches, but I specifically wanted to solve the "send an email to an AI Agent" sort of concept, and give that agent rails to talk back the the human. Doesn't `import email` require SMTP, DNS, signing infra, etc? can it set up MX infra and receive payloads from a mail exchange?

this is interesting. I will have to think more about how humanlayer can support an MCP integration/wrapper, it's not immediately obvious to me

i do think that AI-calling-tools is insufficient to provide bidirectional communication rails for user input/review though...not disagreeing just maybe thinking out loud a little here

helpful - thanks! I have played with temporal a bit but have this thought that since most AI tools represent state as just a rolling context window, maybe you don't have to serialize and entire call stack and you can cut a bunch of corners.

but we're all probably better off not investing that wheel

i mean control loops are good but you don't need hundreds of them

hah yeah I don't know how soon we will be on great accuracy for the latter, for things like "send an email", people tend to just block everything for approval, because clicking approve 90 times hand editing 10 times is a lot better than copying 90 things from one app to another and then 10 things copy, hand edit, send

although I do have some ideas on how you could use vector similarity against past executions to get a 1-100 score on how likely a given action is to be approved rejected. You could set a dial to "anything below 60 just auto-reject it and provide the past feedback to the model preemptively". This would need a lot of experimentation, might even be a research angle (if it hasn't been tried already)

(thinking like cosine * {1 if approved, -1 if rejected} and normalize the score 1-100. You could maybe even weight rejection in 0 to -1 based on sentiment)

thanks - fixed! emailed you separately

yeah it's an interesting point. I can only guess that we didn't do a good enough job of learning from the humans while they were doing their jobs...seems like traditional ML or even LLM tech might be good enough that we can take another pass? Overall the thesis of humanlayer is that you should do all this super gradually, move the needle from 1% AI to 99%+, and have strong SLOs/SLAs around when you pause that needle moving because quality took a hit.

very cool - come ping us in discord happy to help out - we did do a demo w/ dendrite/stagehand a few weeks back where the AI can pull you into a browserbase session OR just ping you in plaintext to get things like MFA codes etc

I think a lot about the low code side and how we can make that work...at the end of the day that looks like a feature/integration into other platforms and that means a lot of matching opinions/interfaces.

I think K8s ecosystem did this well but it required big cross-enterprise working groups that produced things like CSI, SMI, OCI, and before that could happen, there was like 5+ years of storming and forming that had to happen before the dust settled enough for enteprises to step forward and embrace k8s/containers as the new compute paradigm

maybe i'm overthinking things.

onto the coding things -

> clever architecture still tends to beat clever coding

love this

> best code often is the code that’s not written and not maintained

it's too bad not-writing-code isn't as satisfying as deleting code

> it’s developments 5th or 6th version creating major overhauls

yeah the best agent orchestration architecture I'm aware is on interation 4 going on 5. I told him to open source it but he said "its .NET so if I open source it, nobody will care" XD

> There is agentic ai and human in the loop to still figure itself out, as well as how to improve the existing processes. 2025 looks to be interesting.

i'm stoked for it

yeah this is cool. I saw a couple other people posting about this idea. I know some other folks working on "sourcing the humans" or doing a marketplace style thing. Thoughts on things like Payman or Protegee?

thanks - great catch. cool service :)

thanks!

ah okay - I guess in that case, I would like chain a HitL step as an MCP server that wraps/chains to another tool that depends on approval?

or is there a cleaner way to do that?

this is correct - I think helping you BYO humans will help you get much better training/labeling than outsourcing anyways, and that's the end vision of all of this - use humans to train agents so someday you might not need human in the loop, and those humans can move on to training/overseeing the next agent/application you're building

ha fair enough - i think there's another comment thread on just being open w/ 10c / call and i wanna try that out

nice man welcome aboard

meant to reply sooner. It's an interesting problem. I'll have to think on this one.

i think the slack side is easy. I think an AI-optimized email communication channel is a long ways off. I spent weeks throwing things at my monitor figuring out reliable ways to wire DNS+SES+SNS+Lambda+Webhooks+API+Datastore+Async Workers so that everything would "just work" with a few lines of humanlayer sdk.

And what we build still only serves a small subset of use cases (e.g. to support attachments there's a whole other layer of MIMEtype management and routing to put things in S3 and permission them properly)