Unique Research · Interview · December 26, 2023 · Author (CN): Tan Qiuping · Guest: Yang Jinsong, Founder & CEO, AutoAgents.ai
In April, Stanford and Google Research published Generative Agents: Interactive Simulacra of Human Behavior — the Smallville “virtual town” of 25 agents living, chatting, and socializing with distinct personalities, backstories, and interests. First author Joon Sung Park stressed that LLMs/ChatGPT matter, but the decisive piece is a generative-agent architecture with memory, perception/action, and planning/reflection modules — computational agents that behave like credible humans with long-horizon consistency. Many practitioners called it an early AGI marker.
That wave pulled veterans into new companies. Yang Jinsong — former DAMO Academy product/commercialization lead and GM of DAMO green-energy AI who shipped the big-data+AI platform “Alibaba Lingjie” — left Big Tech after watching ChatGPT’s leap in natural-language intent understanding (late 2022) and co-founded AutoAgents.ai (未来式智能) for enterprise Agent landing: natural-language + visual-drag platforms so business users can build custom Agents and chase 10× work efficiency.
Yang graduated from Peking University Yuanpei (CS). Besides DAMO he led Feishu AI & search product at ByteDance and worked at AWS. His products twice won WAIC SAIL awards, twice made DAMO’s Top-10 tech trends, and entered World Intelligence Congress / XinhuaNet digital-model cases. Co-founders span Alibaba, Tencent, and Google — a rare China LLM-application team with both product R&D depth and commercialization scars. Unique Research spoke with Yang on Agent practice and foresight.
Intent-native Agent software is a large opportunity
From Turing’s 1936 abstract machine and Colossus programming to Tukey’s 1958 “software,” software has been the main human–machine medium for ~40 years. Developers flatten every feature for discoverability — PPT alone allegedly has 2,000+ features — while software never knew what humans truly wanted. The gap is chronic over-/under-use. “With generative AI and Agents,” Yang told Unique Research, “a new software form can respond to human intent, raise efficiency and outcomes, break old constraints, and open a great new opportunity.”
Every enterprise will normalize AI tools — and, deeper still, re-iterate bottom business processes with AI-native thinking. Agents are among the best paths for scaled LLM landing. That thesis, plus the team’s prior AI capital, birthed AutoAgents.ai.
Three Agent application classes
(a) Fix human–machine asymmetry. Work that burns human hours on collection and summarization — low cognitive value — is exactly what Agents do well.
(b) Agents as a new traffic entry. Instead of hunting a buried app feature, users speak to an Agent that carries business logic and feels natural.
(c) Multi-agent autonomous collaboration with humans. Highest form: role- and goal-assigned Agents explore an environment together — Smallville is the archetype. Multi-agent dialogue/collaboration is a core research topic because single Agents are capacity-capped, just as humans need orgs. AutoAgents.ai therefore designs Agent org structure, teams, and inter-Agent links that mirror human collaboration.
The company aims to be an “Agent factory” Multi-Agent framework so enterprises can mint their own Agents for cost and efficiency.
Six technical distinctives
1) No-code Agent products for business users; 2) natural-language SOP → auto-generated Agent apps; 3) patented industry-first asynchronous collaborative Agent orgs; 4) first autonomous-agent product proven in live business scenes; 5) Agent-Tuning; 6) growing Agent code-generation ability.
Competitiveness hinges on Agent intelligence and complex-task completion. AutoAgents.ai keeps the “brain” (base LLM) sharp via Agent Tuning and tool-use training; Memory via RAG plus best-practice/SOP knowledge with reward labeling to reinforce good judgments. Raising code-gen and tool packaging from natural language is treated as SOTA work — a paper aimed at next year’s ACL. Async collaborative orgs break serial-only patterns (1:N, N:N, broadcast). Users may only talk to an entry Agent that dispatches task Agents. Enterprises do not start from a blank board: each Agent ships with APIs/tools, embedded best practices and knowledge, plus custom policy overrides.
Two core products
Lingda (灵搭) — LLM app builder for business users: NL goal planning, workflow/tool generation, visual drag-and-drop; 1–2 people can ship an LLM app. Overseas social Autonomous Agents — coach + operator for China go-global / local SMB social accounts: diagnose pain points, auto-build strategy, cycle execution with multi-Agent loops that simulate human social ops.
LLM vendors welcome the stack because models gain real work beyond chat. AutoAgents.ai partners with open-API model providers so enterprise users can call Agent capabilities when needed. Typical architecture: infra (compute/data) → models → mid-layer frameworks like Lingda → apps/verticals. The firm will first standardize energy, finance, government, and marketing solutions, then broaden.
Yang’s guardrails: help clients prepare for human–AI collaboration (a transformation as large as digitalization); avoid thin features a model vendor can absorb in a few iterations by staying close to vertical customer pain regardless of base-model churn.
Closing — remake the enterprise value chain
Citing Marc Andreessen at a16z Connect/Enterprise: generative AI opens new product space and exposes incumbent fragility — a double opportunity for startups. A year ago AutoAgents.ai itself was unimaginable. “Agents are the hands and feet of LLM thought,” Yang said; with reasoning and planning they open a new human–machine era and remake the full enterprise value chain. When Agents go mainstream and best practice is highly SOPified, domain know-how may even be purchasable. The company’s founding aim remains an enterprise LLM + scene platform that puts AI power in every knowledge worker’s hands.
3 CN images; attach on ingest. Author (CN): Tan Qiuping; edit: Qiuping.
— END —