个人助理的下一站：多智能体博弈

Jay Yu 🐟 @0xfishylosopher原推发布于 2026.10.02

在 X 上打开原推

Multiplayer seems to be the next frontier for personal agents. but what might be a path to that happening? Here's my rough mental model: > Phase 1 (where we are today): everyone gets a personal agent. GrokBot, Muse, Instinct etc. one human, one agent (or one human + a team of agents). the agent knows your context, but it still (primarily) interacts with the world the way a human would - scrolling sites, clicking through browser flows, filling forms, calling APIs. > Phase 2: agent-to-agent. At the same time that consumers adopt agents, businesses (doordash, amazon, expedia) are also building in-house agents in their UIs (eg. Doordash's new text-to-order pic). At some point, the agents meet (probs in a matter of months), and A2A happens - my Muse/Instinct simply talks to Doordash/Expedia's agent, and haggles w them to build a personalized price. The interesting part is that negotiation falls out of context asymmetry: each agent knows things about its principal that the other doesn't, and the job is deciding what to reveal. this rhymes way more with OTC trading than with a chat app. > Phase 3: agent swarms and DAOs. Eventually, multiplayer will be more than 1:1 A2A, it will be a "congress" of agents expressing ideas on behalf of their principals. this is where DAOs become interesting again as a way to coordinate. Back in the day i was a DAO voter, but the bottleneck was always attention/apathy - i simply could not read every forum post + proposal. but agents remove the bottleneck of human attention, and can understand your preferences, read every proposal, and advocate on your behalf. DAOs can then fulfill their potential as a structure for spinning up groups to coordinate shared resources/treasuries, becoming a coordination layer for agents to collectively make decisions. To close the analogy, Phase 1 of where we are at is like the PC moment, where everyone has access to agents in the first place. Phase 2 is the Internet moment, where agents start talking to one another, and Phase 3 is where the agent-native, skeumorphic applications really come to life.