Unique Research · Insight · May 30, 2024
With Sam Altman, Andrew Ng, and others “selling” Agents as next-gen productivity, the category finally broke into broader debate. Zhou Jian had already forecast that 2024 would be the year of LLM application landing and Agents.
Why Agents follow LLMs: unsupervised models still lack autonomous action and adaptive goal pursuit in complex environments — capping organizational value. Agents close that gap. So what form do Agents take in ToB, and how do they reshape enterprise software?
Five key Agent traits (Gartner-aligned)
Gartner: autonomous/semi-autonomous software that senses, decides, acts, and pursues goals in digital or physical environments — sensors in, tools/actuators out.
- Adaptability — change behavior/strategy as environment or goals shift.
- Proactivity — anticipate scenarios; act early.
- Goal complexity — manage multiple, often interrelated goals.
- Environment complexity — operate under uncertainty and variation.
- Autonomy — how far with minimal/no human intervention.
Most live apps today are single-Agent. In Lanma’s practice, a GPT-4-class Agent guided by expert knowledge can cover a role held by someone with <5 years’ experience — e.g., drafting regulation-compliant reports from multi-source heterogeneous data.
Inclusive-finance diligence example: bank experts supply a Word template (license, articles, shareholders, bank flows, financials…) plus chapter rules. The Agent pulls APIs / RM dialogue, extracts per business logic, summarizes, and fills — planning, memory, tool use in one loop. With GPT-4o-class real-time/affective models, single Agents will cross more process and role boundaries. Multi-Agent hard parts: value/goal alignment among Agents and with human staff.
What enterprises need
LLMs blow open the feature ceiling of classic software and blur product boundaries — Agents can chain tools to understand and fulfill demand. AI 2.0 “emergence” shows up as: (1) fast information extraction from messy docs/chats (a Prompt engineer’s week vs years of NLP); (2) instruction following in natural language (e.g., “stable job” = each role >18 months over 5 years — no hand-coded rules); (3) adapting to people.
User view: post-ChatGPT everyone expected CUI; GPT-4o argues multimodal real-time. Lanma: Agents should adapt to user expertise — clarifying questions, recommendations so choosing yields the right answer.
Expert/designer view: low-code/BI promised “business users,” but IT still owned the tools. Now Word/Excel can teach Agents; experts may only maintain a Word of knowledge and steps — translation handled by Agent platforms and model vendors. That is the advanced form.
Early enterprise scenarios
Agents as new productive forces: expert knowledge + data + models + compute. Expert knowledge sets value height; data iterates knowledge and generality. Org talent mix and service quality can shift fast.
Bank inclusive-lending diligence: RMs spent huge time collecting/analyzing SME/self-employed files with uneven quality and rigid templates. Lanma’s LLM diligence-report Agent auto-concludes and drafts — saving ~80% of write time, spotting omissions, cutting error rates. (Lanma Tech is a Unique Partner; this piece is inventory-typed insight, not a partner welcome.)
Outlook
If GPT-4o-class capability arrives open-source or near-free from a cloud — what happens to Agents? When scaling laws plateau in 3–5 years? With real-time/emotion understanding, offline sales (insurance, floor staff, wealth RMs) may look like “humanoid robots” directed by a sales brain via phone — who to visit, what to say — with digitized interactions enabling personalized scripts/pricing offline the way e-commerce already does. Experts still iterate strategy in the back office.
Next: stronger reasoning (GPT-5/6-class) + process data → more roles outsourceable like BPO; inside firms, an “enterprise brain” that designs/simulates processes enables one-person super-companies — and a flourishing Agent market.
Charts/photos attach on ingest (8 images).