Accepted Early by Y Combinator : HumanLayer

In partnership with

San Francisco, 2023. Amidst the ever-tightening grip of AI's relentless push for dominance, one man stands alone, crafting a vital defense mechanism that will keep humanity tethered to the digital decisions machines make on our behalf.

This man is Dexter Horthy, founder of HumanLayer, a Y Combinator-backed startup in the newly minted F24 batch.

And the startup? It’s a hard-wired safety net for an AI-crazed world, one where autonomous agents could easily barrel down the wrong path if left unchecked.

At first glance, Horthy’s idea is simple: allow AI agents to perform useful tasks, but not without human oversight when necessary.

Picture an AI bot firing off emails or modifying production databases—high-stakes actions that can’t afford mistakes. HumanLayer is built to stop those bots from running wild without first asking their human overlords for permission. A Slack ping, an email notification—whatever it takes to remind us carbon-based lifeforms that we still have a role in this brave new world.

The Ghost in the Machine: Building AI with a Human-in-the-Loop

Dexter Horthy knows this chaos firsthand.

As an engineer, Horthy has seen what happens when AI agents try to function without human intervention. His previous stints in platform engineering at Sprout Social and container orchestration at Replicated exposed him to the risks involved in trusting AI with sensitive data and decision-making.

When LLMs started generating pseudo-intellectual gibberish, hallucinating details, and missing key knowledge points, Horthy recognized the need for a safeguard.

Synthflow: Build AI voice assistants to manage inbound and outbound calls

Keep your business on 24/7 with genAI. Synthflow’s simple no-code builder lets you set up human-sounding AI voice assistants that can handle call center tasks: real-time appointment booking, lead qualification, handling FAQ, transferring between agents, and more. White label included. Pay as low as $0.08 per minute of conversation. CRM Integrations with Hubspot, Gohighlevel, Zoho, etc. Start for free or let us build your AI receptionist.

Try AI Phone Calling

That safeguard evolved into HumanLayer: a Python toolkit and SDK designed to force AI agents to pause, get human approval, and incorporate feedback before doing anything high-stakes. Whether it’s an AI drafting a client email or manipulating production data, HumanLayer steps in to ask, “Are you sure about this?”

The idea behind HumanLayer sprouted from Horthy's own frustrations.

In a previous experiment, he wanted an LLM to manage customer interactions on a marketing website, but the bot kept veering off course. Missing key pieces of information and sounding like a pretentious teenager, it quickly became clear that this tool could not be trusted unsupervised.

So Horthy wired up a system to get his approval on every important response—this rough prototype eventually turned into HumanLayer.

How HumanLayer Works: Your Last Line of Defense

The core of HumanLayer lies in its approval workflows. You can embed these workflows into your AI tools with just a few lines of code. The magic happens when an AI agent tries to execute a high-stakes action, like sending an email on behalf of a team or updating a CRM record.

At that moment, HumanLayer kicks in, blocking the action and sending an alert to a designated human, usually through Slack or email.

The human can then decide whether to approve or deny the action. If denied, the feedback gets incorporated back into the LLM's context, helping improve future decisions. In other words, HumanLayer doesn't just halt the action—it teaches the AI to do better next time【11†source】.

Key Features: Guardrails for the Digital Frontier

HumanLayer is packed with features designed to tame the unruly beasts that are AI agents. These include:

1. Granular Routing: Route approvals to specific teams or individuals, ensuring the right experts are consulted for the right decisions.

2. Human as Tool: This feature lets AI agents consult humans for feedback, answers, or advice at any time, making the agent’s decision-making process a collaborative effort.

3. OmniChannel Contact: AI agents can seek human approval or input across multiple channels—Slack, Email, Discord, and more—fitting seamlessly into existing workflows.

4. Advanced Approvals: HumanLayer allows teams to configure complex approval flows, including escalations and timeouts, so that nothing slips through the cracks【12†source】.

HumanLayer’s appeal is its versatility. It integrates with all the big frameworks and LLMs—OpenAI, LangChain, and Llama among them—and works in a wide range of use cases, from marketing and customer support to IT and engineering. If a function call has the potential to cause damage, HumanLayer steps in to act as a buffer, ensuring that human intuition and judgment remain part of the equation.

The Stakes: What Could Possibly Go Wrong?

We’ve all seen it in sci-fi: the moment when a machine, unchecked and unmonitored, makes a decision that costs billions—or worse, millions of lives. While today’s AI agents aren’t likely to launch a nuclear strike, they are capable of sending ill-advised emails, mishandling customer data, or making disastrous product updates. These are the kinds of decisions that require more than just machine logic. They need human discretion.

At its core, HumanLayer was built to address the balance between automation and trust. High-stakes decisions are, by definition, too risky to leave in the hands of machines. HumanLayer allows businesses to automate workflows while ensuring that humans are looped in for the decisions that matter most. It's a balancing act between productivity and oversight, efficiency and caution

The Road Ahead: What’s Next for HumanLayer?

The future of AI agents is heading toward complete autonomy, where bots manage entire workflows without human intervention. But we’re not there yet. HumanLayer is playing a critical role in bridging that gap. By introducing layers of human oversight, Horthy’s tool offers a safeguard in an industry that’s racing toward automation without always considering the consequences.

Shameless Pitch Opps.ai

You can also use Opps.ai to find data on your ideal client profile. You can use this data in cold email outreach using our Opps Plus features and then in a tool like Waalaxy or Drippify. Most records have both the ICP prospect’s business email and LinkedIn URL

Use code WolfCast1kFree for a free month of our PRO plan that includes 1,000 free records to find data on your Ideal Client or Ideal Investor. We use this data in cold outreach emails and LinkedIn campaigns.

With ambitious plans on the horizon—including integrating advanced vector databases for smarter knowledge incorporation, and offering enterprise-level features like role-based access control (RBAC) and single sign-on (SSO)—HumanLayer is poised to be a staple for any company that doesn’t fully trust its machines. And let's be honest, who among us does?

In a world where AI is becoming more powerful—and, some might say, more reckless—HumanLayer stands as a crucial line of defense. A little pause before the machines hit “send” might just save us from the next big catastrophe.

So, until the machines can really be trusted, maybe don’t sleep through the next approval request.