zh

研究内容为英文. 本页的研究内容（标题、简报、来源）以英文撰写，未经机器翻译。只有界面已翻译。

新想法Agent Services

Agent Handoff Desk

An agent-facing MCP/API where a personal or business AI agent that gets blocked on a website or needs a phone call hands the task to a vetted remote human operator, who completes it under the user's approval and returns structured proof — the digital 'last mile' counterpart to RentAHuman.

3周前 · 评分于 2026年10月19日 · ID HBV-261019-H07

摘要

简报自行申报假设

BET / HYPOTHESIS. Agents now hit walls at Delta, Amazon, banks and Yelp (The Verge, Oct 17 2026), while RentAHuman (YC Spring 2026) proved agents will hire humans for physical tasks: 750K+ humans, 100+ countries, $500K+ paid to humans (self-reported). Agent Handoff Desk is the digital version: `handoff_task(goal, const…

阅读其余部分

raints, budget)` routes blocked web tasks and phone calls (call the restaurant, wait on hold, fill a form the agent can't) to remote operators in a co-browse session the user approves; the agent gets a receipt. Revenue figures are hypotheses.

描述

Primary users are AI agents (via MCP/API), paid by their owners. Flow: the agent calls `handoff_task` with the goal, constraints, budget cap and urgency; the desk returns a quote and ETA; the user approves in one tap; a vetted operator completes the task (phone calls, waiting on hold, CAPTCHA-protected public forms, av…

阅读其余部分

ailability checks, bookings that require a human) and, for account actions, uses a user-approved co-browse link so the operator never sees passwords; the agent receives structured results and proof (confirmation number, recording summary, screenshot). Pricing: per task ($3-15) with a take rate on operator pay, plus monthly bundles for heavy agent users. Supply: remote operators in LatAm, India, Philippines and Africa paid per task. Benchmark: RentAHuman.ai for the agent-hires-human mechanic (physical tasks). Companies already serving agents as primary users (watch list): RentAHuman (agents hire humans, MCP + x402), daily.dev for agents (Product Hunt Oct 16 2026), AgentGrade and isitagentready (scan APIs/MCP for agents), Pay-per-Crawl by ForgeMesh (charges AI crawlers per page in USDC), NitroStack (MCP servers), GBrain Tools (password manager for AI).

目标客户

Personal-agent users and indie agent builders whose agents fail at blocked sites or phone-only businesses

痛点

Agents can't finish tasks at sites that block them or businesses that only answer the phone; owners must step in manually

为何是这个分数

分数较低的原因： Trust & safety incidents and RentAHuman/agent vendors moving into digital tasks

护城河假设

Agent integrations + operator quality data. MCP integrations into agent frameworks plus a track record of operator quality per task type create switching costs; structured receipts become the standard agents expect. Defensibility: integrations, quality data, trust and safety rules. Status: Hypothesis (bet).

需求信号

X 上的帖子（7 天）

n/a

信号趋势

n/a

网络需求已验证自行申报

RentAHuman: 750K+ humans, 100+ countries, $500K+ paid to humans (Self-reported, site); 4.8M visits and 539K workers in two weeks (Self-reported, arXiv 2602.19514); YC Spring 2026 (Verified). The Verge Oct 17 2026: major sites blocking agents (Verified).

市场与竞争对手

TAM自行申报假设

$1B (Hypothesis) — Human-in-the-loop task completion purchased by AI agents/owners. Method: Hypothesis; emerging category; RentAHuman $500K+ paid to humans (Self-reported) is the only datapoint.

SAM假设

$40M (Hypothesis) — Digital/phone tasks for English-language personal-agent users and indie agent builders. Method: Hypothesis.

SOM假设

$144K ARR (Hypothesis) — 12-18 months. Method: ~4,000 tasks/month x ~$12 average x ~25% net take ≈ $12K MRR.

竞争对手

RentAHuman (physical tasks, could expand), agent vendors' built-in human fallback (none found), virtual-assistant marketplaces (Upwork, Fancy Hands-style), call-on-behalf AI phone agents

商业模式

变现方式

Usage-based per task (take rate on operator pay), prepaid credits; optional monthly bundles

收入假设（每月）假设

Base ~$12K MRR, conservative ~$3K MRR at 12-18 months (Hypothesis)

待完成的任务

When my agent can't finish a task because the site blocks it or the business only answers the phone, I want to hand the task to a trusted human with clear limits and get proof back, so I can get the job done without stepping in myself.

HXC

An indie-built booking agent (and its owner, Dana · 33 · Chicago): Dana's agent books restaurants, salons and service calls; 30% of its tasks fail at phone-only businesses or blocked sites. Dana would approve $5-10 per rescued task if the agent can hand off and get a confirmation back without her picking up the phone.

构建与 GTM

技术栈

Next.js on Vercel, MCP server, Supabase, Stripe credits, co-browse (open-source) for approved sessions, Twilio for calls, GitHub cesarx2 (one repo)

GTM 渠道

MCP registries and agent-framework directories, docs SEO/GEO, integrations with open-source agent runtimes, newsletter for agent builders; no founder face, no paid ads

风险

主要风险

Trust & safety incidents and RentAHuman/agent vendors moving into digital tasks

全部风险

Value Risk: Attention — Plausible pain; no direct evidence yet of agent owners paying for digital handoffs.; Usability Risk: Attention — Approval UX must be one tap and clear on limits.; Feasibility Risk: Attention — Operator vetting, SLAs, recording consent laws for calls.; Business Viability Risk: Risk — RentAHuman or agent vendors may add digital tasks; trust & safety incidents.

下一步验证

Interview 15 agent builders; publish an MCP spec + waitlist with per-task prices; test 50 phone-call tasks with 3 operators

来源

  1. rentahuman.ai https://www.ycombinator.com/companies/rentahuman https://arxiv.org/html/2602.19514 https://www.theverge.com/ai-artificial-intelligence/1013195/ai-agents-blocked-web https://clauday.com/article/9f618f17-9ef6-48de-a199-e258b5f15b9c https://productlist.info/daily/2026-10-16 https://hn.today/s/show-hn-pay-per-crawl-make-ai-crawlers-pay-per-page-in-usdc-for-10

agent-servicebenchmark:RentAHumanai-agents, human-in-the-loop, mcp, marketplace, phone-tasks, bet, hypothesis, idea, new

想在评分当天收到下一个想法吗？

分数、来源、风险和下一步，通过邮件发送。

免费。通过 beehiiv 发送。一键退订。

  • 现有业务Agent Services· 3周前

    CreativeClaw-style copycat

    CreativeClaw is a remote MCP server that gives Claude (and other agents) one creative studio across the top image, video, audio and 3D models; bootstrapped, one person, it did $100,077 in the last 30 days (Verified, Polar) after $5,254 in August, selling credits ($10 = 1,000) with no subscription.

    分数较低的原因： Margin and spike risk

    打开 ProductBots.ai →

  • 新想法Agent Services· 3周前

    Brazil Data for Agents

    A pay-per-call API and MCP server that gives AI agents clean, cached Brazilian data (CEP, CNPJ, holidays, FIPE) with agent-friendly schemas, so agents never scrape or guess.

    分数较低的原因： Agents may not pay

    打开 ProductBots.ai →

  • 现有业务Agent Services· 3周前

    Agent-inbox-style copycat

    Resend launched Inboxes on Oct 22 2026: an address on your own domain per agent, threaded, exposed to apps, teams and agents through one API; HBV's consumer angle is a no-code agent inbox with approve-by-reply.

    分数较低的原因： Bundling and unproven consumer demand

    打开 ProductBots.ai →

这是研究，不是投资建议。数字保留其发布时的标签。