--- title: "Multiplayer AI: Who Writes the Rules?" url: "https://atlan.com/context-and-chaos/issue/multiplayer-ai-who-writes-the-rules/" description: "Four capable agents can still ship a confused launch. The trouble sits in the rules between them, and in whose judgment the next agent can trust." keywords: "Multiplayer AI, AI Agents, Context Engineering, Enterprise AI" --- > Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/ A Context & Chaos deep dive by Rishi Gaurav Bhatnagar, Head of Research at Atlan, a Forbes Asia 30 Under 30 honoree and the author of a children's book. Published October 22, 2026. His argument: multiplayer AI will take many forms, but the handoff stays. Before agents can usefully pass work between them, a team has to settle what those agents may treat as true, who owns those decisions, and how much of that shared context reaches each agent. ## The design gap - **Campaign event example.** The campaign owner decides what the event should be, the website has a different owner, design guidelines belong to someone else, and whoever measures registrations against cost judges whether it worked. Give each of them an agent: the design agent signs off against a standard the website agent has never seen, and the campaign owner's "our best event yet" means something different to the person measuring return. "Every agent has done the job its owner gave it. No one on the team has made a mistake. The system has a design gap." - The gap opens before anything changes hands, because nobody agreed on the goal or on what each agent may know. It becomes visible when work moves and the recipient gets the output without the judgment behind it. - At a multiplayer AI gathering in San Francisco that Rishi helped host, five builders showed different approaches: people and assistants in one space, agents working as a team, finding where that team fails. None settled what happens when work reaches someone with a different goal or right to act. ## Who decides what "best" means - A campaign agent told to make "our best event yet" has to fill in the word: more sign-ups, or the right people signing up. A website agent can check speed and accessibility without knowing the headline promises an unconfirmed speaker. - Shared context needs scope: some decisions belong to the whole team, others to a role or project, and each should reach the agents it applies to without exposing everything their owners know. - Reusable skills raise ownership questions: who updates the skill when the definition of a qualified registration changes, and who decides what can vary for another team. "The knowledge that guides work can be shared; the decision to change it still has an owner." - Diagram: "Same brief, different definitions of best." ## What arrives with the page - The receiving agent needs the audience, the metric, the approved claims, the publishing constraints and the open questions. "Design approving the layout does not mean the website team approved the date." - Rishi's bet: judgment becomes part of the handoff, showing what was checked, against which standard, by whom, and how far the approval extends. - A handoff also needs a route back, so an error traces to the brief or shared skill that supplied it instead of being fixed only at the edge. - Agent protocols move tasks and messages; they do not tell a team which business decisions to trust or whose judgment counts. The starting question: what was done, and what am I allowed to do next? - Diagram: "The output needs the judgment behind it." ## The rules change with the room - Hypothetical cases: two research groups sharing findings under an agreement, and two law firms on opposite sides of an acquisition whose agents must tell a proposed term from an authorized one. - Teammates want the same result, partners guard everything outside one project, and opponents protect different interests, so permissions and acceptance standards differ. A hidden incentive: "If your agent costs you nothing and the seller pays for every transaction it completes, whose interest is it serving?" - Temporary groups leave records behind; what either side may carry into the next project, and who can contest a record of a missed commitment, depends on their agreement. Where trust is thin, the system keeping the record should not also decide consequences; some collaborations will need a referee both sides accept. - Diagram: "Same handoff, different trust." ## Build what sits between the agents - Four teammates with four capable setups get four useful assistants with no way to work together. Four agents have six possible pairs; ten have forty-five. - In early October two open protocols for personal agents arrived: the [Personal Agent Protocol](https://sierra.ai/blog/introducing-personal-agent-protocol) and [PACT](https://decagon.ai/blog/introducing-the-personal-agent-consent-trust-protocol-pact). They add a way for an agent to show who sent it and what that person allowed. Both are early drafts built for a system of trust, a person and a business they chose. Across companies, nothing yet settles how agents share context both sides can rely on. - Memory versus context: "My agent's memory can hold what it knows about my work. Shared context holds what this group has decided, for as long as the right people and agents need it." - Rishi's bet is to govern the context rather than rely on agent memory, which belongs to its owner and shifts with every conversation. How much shared context reaches an agent should depend on who is running the work and what that agent already knows. Context also has to arrive in time for the decision waiting on it. - Without shared context, the team has "four agents moving faster against four different understandings of 'best.' That is chaos, faster." - Diagrams: "Alignment does not scale pair by pair" and "Memory stays with each agent. Context is what the team shares." ## The bets The handoff will outlast today's interfaces and protocols; the context behind it will need owners and boundaries; and as the players change, the rules for sharing and judging work will change with them. No single company or protocol will write those rules alone, and until teams do, agents will keep solving individual problems while the organizational ones wait. Rishi first sketched these questions in [Multiplayer AI](https://x.com/rishigb/status/2099589382818902390) on X in September. Contributions reflect their authors' views; Context & Chaos curates and vets submissions, with no paid promotions. All issues: [https://atlan.com/context-and-chaos/](https://atlan.com/context-and-chaos/).