DeepSeek General Agent Priority

DeepSeek General Agent Priority | Generated by AI

Home 2026.07

Let me search for this to verify what was actually said.Let me pull the fuller transcript section on Agent specifically.Question: What did Liang Wenfeng actually say about agents in the leaked May 20, 2026 investor call — did he say DeepSeek should build a “general agent”?

Answer: Close, but the actual framing is more specific than “general agent” as the headline goal. Two separate parts of the transcript matter:

1. Agent as a rung on the intelligence ladder (not the end goal). He frames AI progress as a staircase: language model → CoT (last year) → Agent (this year) → continuous learning (next, unsolved) → self-iteration singularity → embodied intelligence. He’s explicit that the Agent step will eventually hit a ceiling — after Agent solves everything it can, it still won’t replace employees, because it’s missing continuous learning. So Agent isn’t the destination, it’s one step before the real bottleneck (learning-to-learn).

2. On the product/commercialization question, when directly asked about verticals (finance, medical, legal — pointing at what Anthropic is doing), his answer was:

Go all-out on general Agent first; vertical agents (finance, doctor, etc.) are lower priority. And within that, Coding Agent is the top priority right now because it can accomplish the most and there are many vertical agents built on top of it.

So yes — he explicitly ranks it: general Agent > vertical Agents, and Coding Agent specifically as the current spearhead of the general-agent push, not medical/finance/legal agents.

Also worth noting given your work: he thinks Anthropic’s coding-agent lead is cyclical and won’t hold long-term, and that the real unsolved problem industry-wide (including for DeepSeek) is continuous learning, not agent scaffolding — investors mostly focus on Agent, but researchers are now focused on the learning problem.

One caveat: authenticity is unconfirmed. DeepSeek hasn’t officially verified the transcript, and it’s a translated/auto-transcribed leak — treat specifics (numbers, exact phrasing) as directionally reliable but not verbatim-official.

References:

Back Donate