--- title: "Radical Ventures Partner: Four Major Startup Opportunities in AI Agents" author: "Unique Research" sourcePublication: "Unique Research (FFCAP English library)" originalPublishedAt: "2024-10-08T17:50:00+08:00" canonical: "https://ffcap.cn/en/research/20241008a07b8n00" source: "https://ffcap.cn/en/research/20241008a07b8n00" language: "en" cnSourceUrl: "https://view.inews.qq.com/a/20241008A07B8N00" --- # Radical Ventures Partner: Four Major Startup Opportunities in AI Agents _Unique Research · Insight · October 8, 2024 · Source: Forbes — Rob Toews, Partner, Radical Ventures (Delphina investor); Unique Research Chinese digest_ Sam Altman and Greg Brockman: the next era is not chatting with a single model but systems of multimodal models and tools that act for users — **agents**: AI that pursues open-ended goals, plans long-horizon, uses tools (e.g. browsers), and adapts to new information. **Flight-booking vignette:** check email/calendar → remember seat/time preferences → search/pick flights → fetch PII/payment → buy via airline web/API — no user babysitting each step. Andrew Ng: agentic workflows may drive more AI progress this year than the next foundation model. Andrej Karpathy: AGI will take some form of AI agent. ## Key technical arc Not one paper — stacked advances that wrap structure around LLMs so they can act. Prefer Ng’s _agentic_ (degree) over binary agent/non-agent. - **Chain-of-thought prompting** (Google Brain, 2022) — multi-step decompose/solve; not built for agents or tool use, but core to planning. - **ReAct** (Google Brain, 2022) — first clear join of LLM reason + act; still limited. - **Tool use:** Toolformer (Meta, 2023) — fine-tune when/how to call APIs (calculator, calendar, translate). Follow-ons: **Gorilla** (choose among thousands–millions of APIs); **Chain-of-Abstraction** (multi-step tool plans, outputs feeding inputs). - **Multi-agent:** specialization + modularity beat one mega-agent on hard tasks. Open example **ChatDev** (CEO/CTO/programmer/reviewer/tester roles; arXiv:2307.07924). Frameworks: AutoGen → MetaGPT, LangChain LangGraph. Open questions: hierarchy, shared learning, when to spawn agents, compute as agent count explodes. ## Where to build: apps over infra (for now) Classic pickaxe thesis favors infra first. Many agent infra startups (orchestration, memory, auth, hosting) — but usage still soft even as agent apps proliferate. Space between foundation-model providers and apps for large standalone infra businesses is unclear; many products run tightly coupled internal tools; next models may be agentic natively (GPT-5-class). **Biggest attractive market for agent startups today: the application layer.** **What works:** no reliable general agent yet — constrain to a vertical with structured, repeatable workflows (software eng, SDR, regulatory compliance). Prefer a _natural human-in-the-loop_ already in the workflow (customer support: the customer and/or CS manager). Enterprises spend ~70% of budget on people vs <10% on software — agents are labor, not just productivity tools, and can price nearer to salaries / hiring budgets (leading agent startups already do). ## Four application arenas **1\. Customer support** — contact-center market ~$332B (2023) → $500B+ by 2030. Standardized, repetitive (password resets); natural HITL. Klarna: OpenAI-powered assistant handled ~2/3 of CS requests (2.3M conversations in month one), automated work of ~700 FTEs, ~$40M profit impact cited. **Sierra** — $100M+ from Benchmark & Sequoia; Bret Taylor (ex-Salesforce co-CEO, ex-FB CTO, OpenAI board chair); outcome-based pricing (“software’s future”). Also **Decagon**, **Maven AGI** (A rounds); Maven claims 93% autonomous handle + 60% faster resolve; Decagon customers Eventbrite, Rippling, Substack — differentiation in infra/orchestration around shared base models (Jesse Zhang). **2\. Regulatory compliance** — structured, pattern-heavy, repetitive; analysts + managers as HITL. **Norm AI** ~$40M across two Coatue-led rounds — continuous ops review, flag noncompliance, propose remediation; digests Clean Air Act (~213k words), ACA (~372k), ADA (~22k). **Greenlite AI** — AML/KYC focus; document review + web investigation; replaces offshore contractors with better cost/speed/accuracy/transparency (Will Lawrence). **3\. Data science** — like SWE: complex, high-pay, structured/repeatable. Personalization, demand forecast, recommenders, dynamic pricing, fraud. **Delphina** (ex-Uber DS leaders Jeremy Hermann & Duncan Gilchrist; Radical portfolio) — agents across problem framing → data select/transform → feature eng → train → post-deploy monitor/improve; “junior data scientists” freeing seniors for judgment. **4\. Personal assistants** — sci-fi obvious (JARVIS / Samantha) yet oddly under-built vs other categories; prior waves failed on open-ended daily variability. LLMs + agents may finally make competent assistants reachable — challenge is enough structure without killing flexibility. **Mindy** — “everyone’s private chief of staff”; email-embedded (cc for meetings, forward docs for summaries); 4B+ email users; async enables research before reply + light human review. Seed $6M led by Roelof Botha (Sequoia) & Peter Thiel (Founders Fund) — PayPal Mafia founders. **Ario** — consumer (not enterprise): family calendar, Amazon returns, personalized trip plans; aggregates Instagram, Google Calendar, DoorDash, Fitbit (privacy promised); proactive (e.g. daughter’s birthday + gift ideas from current interests). ## Outlook These four are samples — SWE, yield management, healthcare monitoring, SDRs, product analytics, data engineering, and more. As base models leap, the range of human work agents can take expands. Agents are not just another hype cycle — they are where AI systems are headed; many people already interact with them daily, with stranger transformations ahead. _Full CN ~9k chars + 6 figures — attach on ingest._ — END — --- Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20241008a07b8n00 On-site reading page: https://ffcap.cn/en/research/20241008a07b8n00