HumanLayer
Human-in-the-Loop for AI Agents
HumanLayer is an API that enables AI applications to contact humans for help, feedback, and approvals. Our customer Cased uses HumanLayer to ship DevOps agents that manage complex and risky workflows like production deployments and database migrations. We’ve grown to $500 MRR during the batch. We’re building HumanLayer because we know that the future of AI Applications is not gonna be humans sitting at a chat interface, the future is “outer loop” or “headless” agents, and our partners are building AI apps that invert the typical interaction paradigm. Autonomous agents are calling humans, not the other way around. AI Agents are poised to disrupt the $4.6tn global labor market, but in order to make agents reliable today, and train them to be fully autonomous tomorrow, solutions like HumanLayer are an inevitable part of the AI Agent stack.
Active Founders
Dexter Horthy, Founder
Previously engineer turned product manager at Replicated, a series C dev tools startup. I worked on container orchestrators, helped teams like Hashicorp, DataStax, and H2O.ai deliver and ship their kubernetes-native self-hosted offerings, and continually reinvented the developer experience for Replicated’s core products. I stumbled on the Human-in-the-Loop-for-AI problem while building slack-based agents that managed SQL warehouses. In 3 weeks I pivoted to HumanLayer and launched an MVP
Company Launches
Hi 👋 I’m Dex, and I’m building HumanLayer.
HumanLayer lets your software contact humans for feedback, input, and approvals. It’s perfect for
- Implementing manual approval steps for software- and human-driven workflows
- Oversight of autonomous AI agents
- Managing the transition between the two
HumanLayer has clients for Python and Typescript and lets you manage approvals across Slack, SMS, email, and other channels. For AI agents, HumanLayer is designed to work with any framework + LLM.
You can try it today at humanlayer.dev.
https://www.youtube.com/watch?v=5sbN8rh_S5Q
Ever since I got hands-on with HumanLayer, I’m asking my whole team to think more about Agentic AI workflows - I used to think agents were not worth the effort to get them to production-grade reliability, but now I'm stoked on what is possible when you bring humans into the loop
– Vaibhav Gupta , Founder @ Boundary (YC W23)
Why HumanLayer
Before HumanLayer, I was building autonomous AI agents. Our idea for an MVP was an agent that would coordinate with humans in Slack and could do basic cleanup, like dropping any Snowflake/SQL table that hadn’t been queried in 90+ days. We had already built a platform to generate recommendations, but we wanted to push the limits of agent frameworks, we wanted the agent to be able to actually do the work, not just generate the recommendations.
Of course, we weren’t comfortable with an AI application running raw SQL unsupervised, especially if it had access to drop tables, so we wired in some basic human approval steps.
From there, I realized that the most useful functions we can give to any software are also the most risky. This is especially true for non-deterministic systems driven by LLMs. I realized that anyone who wants to build agents that do meaningful things will need tools for approval and oversight. So, I started hacking on HumanLayer to enable teams to safely build high-impact agents and put them to work ambiently and autonomously in the background.
Use Cases
Whether you’re hard-leaning into AI or not, HumanLayer helps you build impactful automation that you can ship with confidence because you have humans in the loop for the critical operations. Forget months of tuning or broken use cases on every iteration; build confidence incrementally as you observe and test your apps in production (safely!).
- Internal Automations - review/approve outbound emails or automated infrastructure changes
- Dynamic Data Labeling - think RLHF driven by the experts in your company - Use collected human feedback for benchmarking human vs. AI performance and for fine-tuning AI agents
- Ship fast and iterate - Roll out advanced AI features to customers without needing to get to “perfect.” Human review guarantees nothing bad can happen without needing to spend months on evals, tuning, and prompt engineering. Then, iterate based on how your agents behave in production.
- Customer-facing control and visibility - help your customers build trust in automations by giving them hooks to approve/review your app’s actions
Really enjoying playing with @humanlayer_dev - instead of the manual outreach grind, I just hang out in slack and give my agent feedback as it finds new leads
– Tom Granot, founder, https://syntaxcinema.dev
Quick Example
Integrate HumanLayer into your python or typescript app using a decorator or webhooks.
To let AI apps contact a human for natural language feedback/input, just pass the human_as_tool() function into an AI agent’s tool set.
Team
I’m Dex. I am obsessed with helping devs everywhere deliver safe + impactful AI agents.
I spent the last 7 years at https://replicated.com where I helped great developer teams like DataStax and H2O.ai build and launch the kubernetes-native versions of their self-hosted products. I worked on container orchestrators, led product teams, and founded the GTM org.
I can’t wait to build the future of AI Agents with y’all.
Hear from the founders
How did your company get started? (i.e., How did the founders meet? How did you come up with the idea? How did you decide to be a founder?)
Before HumanLayer, we were building autonomous agents that lived in slack. Our idea for an MVP was an AI agent that would drop a Snowflake/SQL table if it hadn’t been queried in 90+ days. Eventually it would create new optimized tables based on past user activity. We had built a platform to generate recommendations, but we wanted to push the limits of agent frameworks, we wanted the agent to be able to do the work, not just generate the recommendations.Of course, we weren’t comfortable with an AI application running raw SQL unsupervised, especially if it had access to drop tables. We didn’t have the time or the resources to spend the 3-6+ months to get it to the ~99.99% reliability we thought it would need in order for us to be comfortable shipping it to customers. So I put on my startup hat, did things that didn’t scale, and wired up a tool-call interceptor to ping me in slack whenever a high-stakes thing was happening.This eventually evolved into a 3 step approval process where someone on our team would approve an action, then the buyer / head of data would approve the messaging being sent, then a message would get sent to a stakeholder or executive like> **Hey, we noticed you haven’t used this table in 90 days - okay to drop it?**Zooming out, I realized that there is a lot of potential in systems that can coordinate human approvals across a team, because they can enable you to ship AI Agents that do really big, meaningful things, safely. The most useful functions we can give to an LLM are also the most risky. We can all imagine the value of an AI Database Administrator that constantly tunes and refactors our SQL database, but most teams wouldn't give an LLM access to run arbitrary SQL statements against a production database (heck, we mostly don't even let humans do that). That is:> Even with state-of-the-art agentic reasoning and prompt routing, LLMs are not sufficiently reliable to be given access to high-stakes functions without human oversight
What is your long-term vision? If you truly succeed, what will be different about the world?
Between require_approval and human_as_tool, HumanLayer is built to empower the next generation of AI agents - Autonomous Agents, but it's just one piece of that puzzle. To clarify "next generation", we can summarize briefly the history of LLM applications.* Gen 1: Chat - human-initiated question / response interface* Gen 2: Agentic Assistants - frameworks drive prompt routing, tool calling, chain of thought, and context window management to get much more reliability and functionality. Most workflows are initiated by humans in single-shot "here's a task, go do it" or rolling chat interfaces.* Gen 3: Autonomous Agents - no longer human initiated, agents will live in the "outer loop" driving toward their goals using various tools and functions. Human/Agent communication is Agent-initiated rather than human-initiated.There are a lot of other problems that I think need to be tackled in this space, including* Toolkits to enable AI Agents to track and self-manage their token spend, sleeping/pausing until a given time/event.* Toolkits to enable AI Agents to extend themselves with coding tools acquire new tools and sub-agents through tools registries* Orchestrators that are designed for Agents that run in the cloud and that can handle asynchronous tool calls. Providing a great developer experience for durable execution in the face of tool calls that require a human response, which ay take hours or days to complete.