Context

NLW argues that despite a year of rapid agent adoption, AI has so far only addressed the individual half of knowledge work, personal chats, personal agents, personal chief-of-staff bots, while most real work happens in shared team contexts: meetings, coordination, shared documents, and collaborative problem solving. Drawing on recent moves from Anthropic (Claude Tag), the OpenClaw maintainers, and the Every team, he argues the next frontier is "multiplayer AI," agents that live in shared team spaces rather than individual silos, and introduces a four-week team framework for getting there. This matters directly to PMs, who spend much of their work at the intersection of teams, coordinating engineering, design, and stakeholders.

The Big Idea

Agents have so far only solved the individually-owned half of knowledge work; the next real leap is moving agents into the shared spaces where teams actually collaborate, with team-owned context, visible in-progress work, and live handoffs, rather than each person running a private agent that only produces a final, unshareable output.

NLW cites survey data showing 42-60% of the workday is spent on collaborative or coordination-heavy work (meetings, communication, shared search), a portion of work current agentic tools barely touch, since almost all agent experimentation so far has been individuals building personal agent teams that serve only themselves.

Key Insights

1. Most of the workday is collaborative, and agents haven't touched that part yet

NLW cites three survey data points to establish the scale of the gap: one study of roughly 16,500 office workers found 42% of time spent working with others versus 39% working alone; another found 57% of time goes to communication (meetings, email, chat) versus 43% to individual creation; a third found roughly 60% of time goes to "work about work," coordination, search, and process. Despite this, he observes that virtually every agentic experiment discussed in the AI community so far, personal researcher agents, personal coding agents, personal chief-of-staff agents, serves only the individual who built it, leaving the majority of actual work time essentially untouched by agent tooling.

2. Claude Tag is the clearest current product expression of multiplayer AI

• What: Anthropic's Claude Tag lives inside Slack channels as a single shared agent instance per channel, rather than each person's individual Claude being pulled into a conversation. Anyone in the channel can see what it's working on, pick up a conversation where someone else left off, and the agent accumulates context about the channel's ongoing work over time rather than starting from scratch with each new person.
• Why it matters: this eliminates the repeated-context problem where every individual user has to re-explain background information to their own private agent instance.
• Example: Anthropic reported that 65% of their own product team's code was generated through their internal shared Claude Tag implementation, not through individual developers using personal Claude agents to submit separate pull requests, evidence the shared-agent model can become the dominant way work actually gets done, not just a convenience layer.

3. Shared visibility during the work, not just shared access to the output, is what makes multiplayer AI different

The OpenClaw maintainers initially tried collaborating on their major 2.0 rebuild using individual agents coordinating through Discord messages, but found this insufficient and built a dedicated multiplayer web interface instead. One maintainer, Colin, described the actual unlock: not simply seeing that a teammate was online, but being able to open the exact same live session and see the same context in real time, without screenshots, copied transcripts, or "here's what happened so far" summaries. This distinction matters because a private agent session that only produces a shareable final transcript still forces teammates to reconstruct context after the fact; a genuinely shared live session lets a second person jump in, redirect, or take over mid-task.

4. "Multiplayer" tools have historically beaten "single-player" incumbents, and AI hasn't had that moment yet

Y Combinator partner Aaron Epstein, quoted in the episode's discussion of YC's fall 2026 request for startups, argues the best work tools of the last two decades won specifically by going multiplayer: Google Docs beat Microsoft Word, and Figma beat Photoshop, both by turning solo creation tools into places teams could work together in real time. His argument, which NLW endorses, is that current AI agent tools are stuck in the pre-multiplayer phase Word and Photoshop once occupied: you open a chat, get an answer in a box only you can see, and the best available option for collaboration is sending a read-only transcript link.

Mental Models & Frameworks

Four dimensions for scoring whether a shared work stream is ready for a shared agent

• Shared need: how many people on the team need the same context for this work; more people sharing the need makes it a stronger candidate.
• Staleness cost: how much it hurts when each person's individual version of this work drifts out of sync with everyone else's.
• Permission sensitivity: how much restricted or sensitive data the work touches, which raises the bar for how carefully a shared agent needs to be governed.
• Checkability: how quickly and reliably someone can tell whether the agent got the output right.

NLW suggests scoring each recurring team work stream on all four dimensions using a one-to-five scale, then prioritizing the highest-scoring candidates as where to pilot a first shared agent. This framework is useful beyond the specific sprint it comes from: it gives teams a structured way to decide which of many overlapping work streams is actually ready for a shared agent versus which should stay individual for now.

The four-part shift from single-player to multiplayer AI

• Private outputs to visible work: teammates go from seeing only a finished answer to being able to watch the agent's in-progress work as it happens.
• Feedback prompts to live participation: teammates go from only being able to react after output already exists to being able to redirect, annotate, or join mid-task.
• Personal memory to shared context: durable context stops belonging to one individual's account and instead belongs to the team, channel, or project.
• Individual leverage to team capability: agents shift from being personal efficiency tools to becoming reusable organizational infrastructure that outlives any one person's usage of them.

Practical Application

Inventory what your team is actually doing with AI before building anything shared

Before introducing a shared agent, take stock of where every person on the team currently stands: who is still just prompting ChatGPT or Claude directly, who has built a recurring personal agent, who has started using context files or skills. NLW notes teams almost always find a wide spread here, and understanding that spread is a necessary first step before attempting a team-wide shared-agent rollout, since people are starting from very different baselines.

Map your team's overlapping work streams before deciding where to deploy a shared agent

For each recurring work stream your team handles, identify what it is, who else touches it, what context it requires, and how often it runs. Work streams where multiple people's individual work genuinely overlaps are the real candidates for a shared agent; work that stays cleanly individual, even if AI-assisted, likely doesn't need this treatment yet.

Pilot exactly one shared agent, on real work, with at least two people using it

Rather than rolling out shared agents broadly, pick a single candidate work stream (ideally one that scored well on shared need, staleness cost, permission sensitivity, and checkability), set it up as a shared agent using whatever tooling is already available (Claude Tag being the most direct current option), and have at least two team members actually use it on real work for a defined period. If it doesn't clearly improve the work, move to the next candidate rather than assuming the shared-agent approach itself has failed.

Questions to Consider

• Looking at our team's actual work, which recurring tasks genuinely have multiple people needing the same up-to-date context, and are we currently forcing each person to maintain a separate, drifting private version of that context?
• If a teammate needed to jump into an agent session we're running mid-task, could they actually see what the agent is doing right now, or would we have to stop, summarize, and hand off manually first?
• Which of our team's collaborative work streams would score highest on shared need, staleness cost, permission sensitivity, and checkability, and have we actually piloted a shared agent against the strongest candidate, or only against whichever use case happened to come up first?
• Are we assuming that because individual team members are already effective with personal AI agents, our team as a whole is therefore AI-native, even though most of our actual work time is spent on the collaborative work those personal agents don't touch?

Bottom Line

Individual AI agents have proven themselves, but they only address roughly half of real knowledge work, the solo half. The next competitive advantage will come from moving agents into the shared spaces where teams actually collaborate: team-owned context instead of duplicated personal folders, visible in-progress work instead of private outputs, and live handoffs instead of after-the-fact transcripts, the same multiplayer shift that made Google Docs beat Word and Figma beat Photoshop.

Concepts to Explore

The single-player to multiplayer pattern in software history

NLW and the episode's cited YC partner point to Google Docs displacing Microsoft Word and Figma displacing Photoshop as the two clearest historical examples of a category winning specifically by becoming multiplayer, turning what had been an individual creation tool into a shared, real-time collaborative space. This pattern is worth studying for anyone building or evaluating AI tooling, since the episode's central claim is that AI agents are currently sitting exactly where word processors and design tools sat before their own multiplayer transitions.