humanlayer

Building the post-IDE IDE @ THAT WORKS @

[humanlayer @humanlayer_dev]

👀

[dex @dexhorthy]

superthread of every resource I have published about advanced claude code usage, research-plan-implement, and advanced context engineering for coding agents 👇

(mostly chronological)

[Raduan Al-Shedivat @0xRaduan]

masterclass from @humanlayer_dev

[Skylar Payne @skylar_b_payne]

This example uses @humanlayer_dev which you should all check out -- great library for human in the loop workflows!

[Bilgin Ibryam @bibryam]

🧑‍💻 @humanlayer_dev

Best expressiveness and dev experience I’ve seen.

You can annotate any tool for human approval.

But it adds another layer to your agent stack. If you’re good with that, it’s absolutely worth it.

https://t.co/Mi85D9c1PP

[Rrahul @rahulx004]

#100DaysOfLearning
Day 7/100

12-Factor Agents:
Patterns of reliable LLM applications — 
@dexhorthy , @humanlayer_dev

https://t.co/QK6Xbf7zJO

now i have to actualy write some code also on @freeCodeCamp , lets codeup the legacy javascript course here #noAIuse

[Nayak Satya @NayakSatya_SG]

5 APIs you will love for AI integrations and automated workflows smoothly

1-: Bluesky Firehose API

If you want to use social media posts to analyze public sentiment about a particular topic then Bluesky firehose api is best fits.

It is already being used for hacking projects like AI training.

2-: Signature API

If you need authentication and certification then this is your go to API

3-: Seam API for IOT

It is a universal system for controlling a matrix of devices for the Internet of Things.

4-: HumanLayer API

HumanLayer API provides the structure and integration format for AI to seek human contact when they need it.

It inverts this paradigm with an API framework that allows computers to contact humans.

5-: Hugging Face Transformers API

Hugging Face’s Transformers API makes it relatively easy to fire up PyTorch, TensorFlow or JAX and access dozens of foundation models.

Video credit - @humanlayer_dev

[Alex Reibman 🖇️ @AlexReibman]

4/ HumanLayer 🏅Most Practical

LLM powered email classifier that dynamically updates detection rules based on human the loop notifications over Slack

Human powered guardrails for AI

@dexhorthy @humanlayer_dev

[dex @dexhorthy]

our team has been hustling super hard ever since the claude code sdk dropped - saw this tweet and couldn't help but show off some of the very dope work they've put in - early days but already looking very cool

[dex @dexhorthy]

Claude Code SDK is a PERFECT example of 12-factor agents in action. They handed us the keys to the loop instead of having it be a black box owned by the `claude` CLI

Now all the best hackers I know are building their own bash pipelines and custom loops, dropping in `claude -p` as a focused micro agent, and owning the "outer loop" between the agent and the human

much more cool stuff coming soon here from @humanlayer_dev - stay tuned

[dex @dexhorthy]

If this tweet gets 1000 rts @pk_iv will grow a mustache for ai engineer worlds fair

[humanlayer @humanlayer_dev]

@0xblacklight @browser_use very dope use case! thanks for sharing!

[Kyle Mistele 🏴‍☠️ @0xblacklight]

Built an agent that logs into my twitter account with @browser_use + @humanlayer_dev so it can ask me for my password & a fresh 2FA code when needed.

Bullish on systems that treat humans as peripherals for agents to accomplish things agents can't do alone (e.g. get a 2FA code)

[anu @sigdel29]

Also S/O to @humanlayer_dev -if you’re building human in the loop AI agents , definitely check them out!

[Joe Heitzeberg @jheitzeb]

I thought I was kidding, right? “Guys, dinner!!!” @dexhorthy @hellovai @boundaryML @humanlayer_dev

[humanlayer @humanlayer_dev]

@LandseerEnga @mignano this guy gets it

[dex @dexhorthy]

everyone loves reasoning models...but did you know with a little prompt pipelining, you can get 4o-mini to perform in the same league as some beefy thinking models?

come hang with me and @hellovai tomorrow to see it in action!

[humanlayer @humanlayer_dev]

d) i'm in this picture and I like it

[dex @dexhorthy]

agentic ai is real. our inbox drafts replies, pings us in slack, learns from feedback, and only bugs us when needed. no busywork, just focused connection. diagram here: 👇

[Vaibhav Gupta @vaibcode]

over 300 ai engineers came to our 3 hour workshops. They asked for another one.

next 4 weeks, 1 hour every tuesday:​🦄 ai that works
First topic: large scale classification

@dexhorthy @boundaryML

[Denied @denieddotdev]

AI Agents with Parental Control using Eunomia + HumanLayer 👮‍♀️ 🚸 🚨

We're kicking off our authorization examples series with built-in parental control for any agent your children might interact with.

LLM-powered products designed for learning, play, and interaction with children are on the rise. Discussing history with a virtual teacher who answers any question is certainly more engaging than sitting through three hours of pre-made slides.

However, allowing children unrestricted interaction with AI agents has limitations. In child-focused AI products, authorizations are often necessary, and Eunomia is designed to help.

To enhance this, we integrated human checks, allowing parents to decide whether their children can access specific content. We used @humanlayer_dev for this.

Find the example link in the comments, built with @langchain , web search via @tavilyai and authorization powered by Eunomia from @whataboutyou_ai

[dex @dexhorthy]

shipping AI agents to the enterprise? here's 3 challenges I'm seeing the most in working with AI startups

1. model data boundaries - many enterprises wont send data outside their environment to hosted models. That means you wanna support BYO models in AWS/Azure, or bundle an OSS model like llama in your app

2. app data boundaries - similarly, enterprises need control over the data that agents store between llm calls - memory, context, rag, etc - that means shipping your app into their datacenter (be it AWS, Azure, GCP, on-prem) - VM images, docker compose, are common here, but k8s tends to be preferred by ENTs

3. oversight and compliance - if your agent is going to do anything useful, you need to prove that critical actions are DETERMINISTICALLY reviewed by humans, and be able to surface audit trails to verify it