At @decasonic, for the past month, we’ve been exploring what comes after AI systems scale to the mainstream... On this horizon, multiplayer AI interfaces that bring people and agents into shared workflows and improve through reinforcement learning. Most AI today is still single-player: one person, one model, and one chat session at a time. Increasingly, we're seeing that chat interface disappear into native iMessage interfaces presented by Muse and other personal AI agents. When the session ends, much of the learning generated through that interaction disappears with it. If there were improvements, they weren't shared across the firm, transparently to compound the expertise. My colleague @abdulalali has been trailblazing the frontier of these RL interfaces, anticipating what's next for consumer AI and personal assistants. What we've been cooking: AI interfaces live where decisions already happen across chat, calendars, meetings, notifications, ranked feeds, voice, and generative UI. When upgraded, these interfaces can learn from explicit signals such as approvals, corrections, reactions, rankings, and preference changes. They can also learn from implicit signals such as accepted suggestions, dismissed notifications, completed actions, meeting outcomes, and changes in workflows. But reinforcement isn’t universal. It depends on context. A dismissed notification could mean “wrong answer,” “wrong time,” “wrong device,” or simply “not now.” An edited response could signal a factual correction, a stylistic preference, or a change in the user’s objective. The interface isn’t just how we access AI. It is part of the learning system. That requires a strong control plane with clear permissions, approval boundaries, confidence indicators, history, corrections, and an RL ledger. People should be able to see what the system has learned and have the power to change it. Our work started with single-player AI through a CLI. The next step is bringing reinforcement learning into multiplayer AI interfaces embedded natively across shared workflows. When many people contribute feedback, corrections, decisions, and outcomes, the learning loop becomes broader and faster. Each interaction can improve the system. Each workflow can generate reinforcement. Each human can contribute to shared expertise, shared intelligence across the entire firm. That’s how intelligence begins to compound across the AI-Native firm. The future of the AI Native firm is a network of multiplayer AI interfaces, upgraded through reinforcement learning, that transforms everyday work into a collaborative learning system.

We’re watching the application layer of agentic finance form in real time. Arc already has meaningful concentration across agentic banking, trading, and payments. As agents gain the ability to transact, allocate, and coordinate capital, the financial rails underneath them become an increasingly important part of the stack.

As connector ecosystems expand, the value of an assistant will increasingly come from how well it can understand context, move between services, and complete workflows with the right permissions.

Personal AI is finding consumers quickly. Muse reached 2.8M installs in its first 12 days. On comparable U.S. and Canada iOS data, it reached 1.8M downloads versus 1.3M for ChatGPT at the same point after launch. Early downloads are only the first test. The more important measure will be whether consumers repeatedly hand these products real tasks, context, and permission to act.

what's truly impressive to see about @Muse is the level of trust. you are able to see the tool calls, the interactions initiated by the agent, and a live view of the browser session when in usage. a live view of what the agent is initiating step by step. great demonstration of generative UI to accelerate adoption and win customer trust.

Safety is becoming part of the intelligence stack. As robots leave fixed cells, they need real-time perception and control around people. Digit 5’s close-proximity safety architecture is a good example of why capability and safe deployment have to improve together.

Physical AI is moving from programmed behavior toward learned behavior. As systems need fewer task-specific instructions and can generalize from demonstrations, the economics of training and deployment begin to change. The value increasingly moves into the models, data and feedback loops that make that adaptation possible.

Arc mainnet going live feels like a real tipping point for agentic finance. Most “financial agents” have still been operating inside human financial infrastructure. Give an agent access to an account, allocate capital, set guardrails and let it execute. Arc appears to be pushing toward a different model: financial infra built with machines and humans An agent can have a policy-controlled wallet, make usdc payments, discover and pay for services. Arc Studio even pushes the builder side toward agents being able to create applications. More importantly, there is an actual financial system underneath it. StableFX brings 24/7 FX (localized payments) BUIDL, USYC, and other tokenized assets create productive collateral. Aave + morpho + uni create open credit and liquidity layers. Agents are not limited to taking actions inside one application they can increasingly interact with a fully composable market. The institutional layer makes this especially interesting. DTCC, Visa, blackrock and others participating as founding network operators is a very different starting point from launching another retail L1 and hoping institutional liquidity eventually arrives. What we are also seeing is that the endpoint may not just be software agents. The next step would be physical machines autonomously negotiating and settling payments on Arc without a human in the loop. Obviously there is still a gap between “Can I” and “Would I”. But the pieces are starting to come together. We @decasonic are actively evaluating the Arc ecosystem and where value ultimately accrues as this market develops.

Agentic finance becomes real when agents start moving capital. @arc offers an early view of the infrastructure behind that shift through programmable wallets, stablecoin payments, deterministic settlement and tokenized assets. The next proof point is sustained autonomous activity across payments, collateral and on-chain markets.

Interesting points here. Canton has been capturing the TradFi side of onchain finance for a while, and the institutional pull seems to keep getting stronger. The part I find most interesting is what comes after more assets actually move onchain. If an asset manager can hold billions of dollars of treasuries onchain, the opportunity starts to expand into how that capital gets traded, financed, used as collateral and moved around an always on 24/7 market. That is also why we’ve increasingly started calling this category fintech at @decasonic . The line between crypto infrastructure and financial infrastructure has blurred pretty heavily. You have the same institutions, assets and capital markets activity, just moving onto new rails with different capabilities. @CantonNetwork feels particularly well positioned here because privacy is still (and will always be) a real requirement for institutions, while composability is what makes these assets more useful and actionable over time. Robinhood is interesting from the other side. Retail is already sitting at the frontier of adoption, experimenting with what tokenized equities can actually do once they are onchain. I expect institutions to catch up to a lot of those use cases as more assets and liquidity move over.

the more interesting takeaway here is that we finally have reached enough history in crypto to separate markets with real underlying demand from products that worked because incentives, liquidity and narrative + market timing were there. Speculation clearly has durable demand. So do stablecoins and RWAs as crypto connects back into the traditional economy. A lot of what sat in between looked like PMF during a bull cycle, but struggled once the incentives disappeared and capital became selective. What stands out to me now is the convergence. Coinbase, Robinhood, MetaMask, Kalshi and others all started from very different places, yet increasingly they’re building around the same identified demand pools. The interesting investment opportunity from here is figuring out what becomes the next durable demand pool before every major distribution platform is building toward it.

Physical AI is moving quickly into the real world. That makes evaluation infrastructure increasingly important. At Decasonic, we believe the next phase of physical AI will require trusted benchmarks that can measure how models and embodied systems perform across real tasks, environments, and hardware. @robocurve is building toward that layer. As the ecosystem expands across models and embodiments, shared benchmarks can create the confidence needed to make deployment decisions and scale adoption. We’re excited to support @chooi_jeq and the Robocurve team following their $10M seed. Congratulations to the entire team 👏👏

Consumer AI is becoming easier to reach in the moments that matter. Gemini on Windows puts the assistant one shortcut away from the user’s existing workflow. That kind of access can become a meaningful driver of habit and repeated use.