Multiplayer AI needs to move past the group chat phase. Chats are chaotic, everyone sees everything (creating social & management issues), and they're bad for long projects. Worse, they keep AI trapped inside existing teams, when its real promise is working & helping across them.
agents need infrastructure for their decisions
humans need infrastructure for their decisions
the future needs decision infrastructure
without it, agent-accelerated organizations descend into chaos
@yoheinakajima A decision that was right last quarter can be wrong today, so I think the playbook needs to show when a rule has been superseded. Otherwise a perfectly retrieved answer can still tell the agent to do the wrong thing.
🚨fresh podcast out!
fun dive into the history of decision infrastructure, how to coordinate effectively across teams of humans and agents, and of course @primitive_hq with @heyitsnoah
this is definitely better than nothing, but almost every design team I meet with is doing this and it isn't working
good news is that a design system is essentially a collection of decisions systematically enforced across product development over time...
like @primitive_hq
it’s true—and bigger models won’t fix it.
given the way agents work today (loops inside loops inside loops), compounding drift is a huge problem that gets worse the more loops you add
imagine you’re a healthcare CEO who wants a report on the state of your business. you kick off a fleet of agents, who spawn subagents, crawling your enterprise knowledge base, mcp connections, etc etc
but early on, deep in one of those subagent reasoning traces, it does something smart! it uses a metaphor: “nurses and patients are like plants and animals”
in context, it makes sense and contributes to the agent’s understanding
but now those tokens are a contaminate that shifts the distribution in latent space for the next loop, and the next loop, and the next loop….
and by the end, the final healthcare enterprise report has a zoology section.
is this example a bit cartoonish? maybe, but if you’ve ever worked with long-running agents you recognize the pattern. it’s why higher reasoning often makes outputs *worse* and bigger models will simply be given longer running tasks which means anything short of perfect (which of course usually has no definition) will compound to the same result.
we’re fixing this drift with decision infrastructure at @primitive_hq. if you’re leading a team struggling with drift, my DMs are open and we have one more design partner spot to fill, let’s get your agents back on track.