Launch HN: Human Layer (YC F24) – Human-in-the-Loop API for AI Systems

Hacker News Discussion ↗

Product concept & motivation

  • HumanLayer offers “human‑in‑the‑loop” (HITL) as an API so AI agents can pause, get human approval/input via channels like Slack/email, then resume.
  • Many commenters say they’ve built ad‑hoc versions for internal workflows and see this as a real, recurring need.
  • The goal is to make agent adoption safer and more controllable, especially where autonomous actions are risky (payments, external emails, operations).

Async / outer-loop orchestration

  • A major pain point discussed: current agent frameworks don’t handle long‑running or asynchronous tool calls well (e.g., waiting hours/days for a human).
  • Several people describe solutions using Temporal, DBOS, MCP, or custom workflows that:
    • Fire async requests
    • Persist state/context
    • Resume workflows on webhooks or signals.
  • There’s debate on whether a “rolling context window” is enough vs. richer, domain‑specific state machines.

Integrations & competing tools

  • Comparisons made to Temporal, Anthropic’s Model Context Protocol, LangGraph/LangChain/CrewAI HITL, Make.com’s beta HITL, n8n/Zapier/IFTTT, Slack/email bots, and review-form tools.
  • Some argue generic automation tools already cover simple “send for approval, then continue” flows; others say HumanLayer’s routing, escalations, multi‑channel support, and observability add significant value.

Pricing & business model

  • Current framing (~$0.10 per operation, $20/200 ops) triggers strong price sensitivity for smaller startups.
  • Concerns:
    • High marginal cost compared to cheap LLM calls and DIY serverless workflows.
    • Free tier vs. paid tier per‑op cost inconsistency.
  • Suggestions:
    • Simpler “$ per action” pricing with volume discounts.
    • More generous starter credits; potentially open‑sourcing core backend.

Use cases discussed

  • Back‑office automations, ops/finance approvals, external‑facing communications, sales emails, payments, LinkedIn outreach, MFA and CAPTCHA‑like “pull a human into a web session,” and agent oversight for web‑browsing bots.
  • Many emphasize they are unwilling to “hand the wheel” to agents without human checkpoints.

Human factors, risks, and ethics

  • Concerns about:
    • Automation bias and complacency: humans may rubber‑stamp approvals once they trust the agent.
    • Decision fatigue and ownership dilution when too many approvals are required.
    • Potential for exploitative outsourcing if a future product version ever “provides” humans.
  • Mitigations proposed:
    • Strong attention‑activating confirmations (typing repo/table names, “signed‑off by” fields).
    • Undo windows vs. hard confirms, with trade‑offs in stress vs. safety.
    • Learning from past approvals/rejections to adapt which actions need strict review.

Technical implementation notes

  • Email support is seen as non‑trivial: DNS/MX/SES/SNS/Lambda/webhooks, MIME/attachments, storage, and routing across many conversations and agents.
  • Slack is considered simpler, but scaling to many users, orgs, timeouts, and escalation rules still adds complexity.
  • Some argue a simple script plus basic SMTP is enough for small cases; others say production‑grade reliability and async orchestration justify a dedicated service.