Looking Forward — What's Next
6 · Looking Forward — Where This Goes
Part 6 of 6 · ~6 min read
We’ve covered the idea, the architecture, the two running apps, and the build story. The last question is the most interesting one: where does this go?
Not in the science-fiction sense. In the very concrete next-12-months sense. What’s plausibly in the cards for ResearchOS, for AI Institute, and for the broader pattern of orchestrated AI agents running on cheap infrastructure?
Three concentric circles, smallest first.
Circle 1 — What ResearchOS becomes next
Near-term, the platform itself has a clear set of next steps:
A third application
Right now there are two apps. The main thing missing is a third — ideally one that’s not research-shaped. Maybe code review across a private repository. Maybe automated journalism (sourcing, drafting, fact-checking). Maybe legal-document review.
The reason a third app matters isn’t capacity. It’s abstraction testing. The two existing apps share a lot of shape (both ask AI workers to produce text grounded in sources). A genuinely different third workload is what reveals which parts of the platform are well- factored and which were accidentally specialized.
I have ideas for the third app. I haven’t picked one yet.
Per-tenant isolation
Today the platform supports multiple “client apps” with proper ownership scoping (each app sees only its own workflows, sessions, ideas). What’s missing for genuine multi-tenancy is per-tenant rate limits and quotas. Without those, one app can starve the others. With them, the platform could plausibly host friendly third-party apps, not just my own.
Smarter recovery
The platform has a watchdog that catches stuck tasks and a node- liveness check that catches dead workers. What’s missing is graceful degradation — a clean handoff when an entire model provider has an outage. Right now I find out from manual monitoring; ideally the system finds out and routes around the outage on its own.
A workbench you’d actually want to use daily
The current workbench dashboard does the basics well — kanban for ideas, summary tiles for projects and recent activity, deep-link to research sessions. What it doesn’t do yet is delight. Drag-and-drop between lanes. Inline rendering of report previews. A “convert this idea into a workflow” button. These are weekend-project features but they’d make the workbench something I’d want to live in.
Circle 2 — What AI Institute becomes next
If ResearchOS is the platform, AI Institute is the showcase application. Three plausible directions for it:
Memory across days
Right now each day’s briefing starts fresh — the analysts don’t remember what they wrote yesterday. Adding a real persistent memory (not just file storage; narrative continuity) would make the briefings sharper. “The energy analyst noted yesterday that crude broke $80; today’s update should explicitly comment on whether that held.” That kind of cross-day reasoning is currently missing and is the obvious next leverage point.
Real-time data integration
Today the analysts work from whatever the underlying AI models can search the web for. Wiring in proper data feeds — Bloomberg-style market data, news APIs, regulatory filings — would dramatically upgrade what’s possible. The platform supports this; I haven’t built it yet because the data licensing is a separate problem.
A small group of beta users
If the operation continues to feel useful for me personally, the obvious next step is opening it to a small group (say, 10) of investment professionals as a beta. Not as a commercial product — as a way to find out which parts of the workflow generalize and which are idiosyncratic to me. The platform’s per-app isolation makes this feasible technically; what’s not yet figured out is the licensing of the underlying AI usage and the legal posture of an “AI investment research” service that’s not registered investment advice.
That’s not a small project. It’s the project I’d start if I were to commit to AI Institute as a full-time thing rather than a side exploration.
Circle 3 — What this pattern means more broadly
The most interesting question isn’t what happens to my project. It’s what happens to the broader pattern.
The pattern: **one person + AI partner + cheap edge infrastructure
- a few specialized AI workers = a small but real production system that does meaningful work.**
That pattern is reproducible. It’s not a stunt. It’s now within reach of anyone willing to spend a couple of weeks learning the ropes. And it’s about to enable a class of small-scale software that didn’t really exist before.
A few predictions for the next 12-24 months that I’d put real money on:
1. Vertical AI orchestration platforms become a real category
Right now most “AI platforms” are horizontal — they let you run AI on anything. The interesting layer is vertical: orchestration platforms tuned for a specific industry. AI for investment research. AI for clinical trial coordination. AI for supply-chain analysis. AI for niche legal practice areas. Each one looks like AI Institute but specialized. Each one is buildable by a small team, possibly even by one person.
2. The “individual builder operating at SMB scale” becomes common
The tools have collapsed the cost of building SaaS-grade software to near zero. The people who notice this first will build a generation of small, specialized, profitable software businesses operated by one to three people each. Not unicorns. Just businesses that work, that didn’t exist before because the build cost was too high.
The pattern was already happening with no-code tools. AI partnership makes it happen with full software, not just glue layers.
3. The skill that compounds is “describe what you want clearly”
The bottleneck has moved from production to specification. The people who can articulate what they want — in plain language, with unusual precision, attentive to edge cases — will build dramatically more than people who can’t. This is a skill. It’s trainable. It’s not exclusive to engineers.
If you read this whole series and your only takeaway is “I should practice writing clearer requests,” that’s the most actionable thing to take.
4. AI agents become a layer in everyday business operations
The way email and spreadsheets are layers everyone uses without thinking, AI agents will become. Most people will never build their own platform like ResearchOS. Most people will use platforms that someone else built — for marketing, for finance, for HR, for sales. The pattern in this series is the generic shape of those tools.
The interesting question for any business reader: which of your current human workflows are about to look like AI Institute — where a team of specialized AI workers, coordinated by a thin layer of orchestration, produces 80% of what a human team produces, at 1% of the cost?
If the answer is “none,” I’d encourage you to look harder. If the answer is “a lot,” the next question is what role you want to play in that transition.
A more personal closing
When I started building this on April 18, I didn’t have a grand plan. I was curious about whether one person could plausibly stand up a multi-agent AI platform without a team. I built a prototype to find out. The prototype turned out to work. Then I built a second app on top of it, and that worked too. Then I built the dashboard, the documentation, the roadmap, this article series.
What I have at the end is not a product. It’s a working system, a small body of writing about what was possible to build in ten days, and a strong sense that the gap between “interested in AI” and “actually building useful things with AI” has become much smaller than people generally realize.
That gap is what these six articles are trying to close, in a small way, for you.
If the platform itself becomes more — a product, a service, a business — that’s a separate decision I haven’t made. If the articles are useful as a glimpse of what’s now possible, that’s already the outcome I was hoping for.
Thanks for reading. The next move is yours.
Previous: The Build Story
If anything in this series resonated and you want to talk about it — what you’re building, what you’re thinking about building, how this applies to your domain — I’d love to hear from you.
The articles in this folder are released under CC BY 4.0. Share them. Adapt them. Send them to a friend who’s curious about AI but hasn’t seen what’s now possible.
Singapore · April 2026