- Building
We're building Fund Edition, an AI investment-team app. Our project has two human team members. We used a lead agent to coordinate implementation, research and review agents, and the founder tested the product throughout.
There are two ways into the product.
Build an AI fund. Tell Maya what you want to invest in. She helps turn the idea into a team and a paper fund. Review the proposed colleagues, hire additional specialists and give them jobs in ordinary language. Maya is the main contact, but you can also ask a researcher or risk analyst a specific question. We gave the team an office because the founder wanted hiring and directing people to feel like running a small business.
Explore AI funds. This has two sections. Popular active funds starts from public strategies and disclosures of funds such as ARKK and Vanguard Wellington. Investor-inspired funds starts from publicly described approaches such as Buffett's or Dalio's. Each strategy needs its own explanation of the research, roles, sources and historical method. We want people to be able to ask the team what it is doing instead of trying to interpret a chart alone.
My Portfolio brings the visitor's own funds and selected catalogue strategies into one paper account. The demo uses simulated capital. Historical coverage varies by strategy. Our research teams are independent of the original funds and investors.
The interesting engineering work was connecting the conversation, the team roster and the account. A message about wanting less concentration has to reach the right fund. A decision has to use the right evidence. A proposed trade has to go through the execution engine. We found several places where the screen looked right while the state underneath was wrong.
- Blocked
November 17 — duplicate requests and imaginary fills
The founder brought experience from an earlier personal trading system, including a duplicate-fill problem. We built the execution boundary around that concern. The model proposes a target. Separate code checks permissions, price, cash, quantities and position limits before updating the simulated ledger.
A simple check sent the same synthetic idea to two accounts. At a 12% position cap, one account bought seven shares. At 1%, the budget couldn't cover a whole share and the other account kept cash. Later, eight concurrent copies of one Wellington opening request produced one subscription and one opening receipt. Repeating a quarterly action after revoking sell permission advanced one period and left the positions intact.
November 18 — the public manager went quiet
Maya could hire a colleague locally. On the public site, the greeting played and speech input worked, but a new question never got an answer. We had treated a loaded page as evidence that the app was live.
There were two faults. HarnessRouter declined our hosted intermediary route. The cloud package also had an older studio than the founder had reviewed. Frontend aligned the screens and deep links, backend checked the extracted package, and release moved the application to a direct server with persistent storage.
Railway then rejected the Dockerfile's VOLUME declaration for /data. The platform already managed that mount. Release sent the failing line to packaging. We removed it while keeping a separate office redesign out of that deployment.
Acceptance started from a new public visitor, hired Jules and asked what cash-flow evidence would make the researcher wait. Jules answered. We restarted the service and checked that the same visitor still had the team and conversation. This became a release check we could actually repeat.
November 18 — the founder kept finding the wrong entry point
The founder chatted with Maya for a long time and asked when hiring would happen. We had built a setup conversation without a clear transition to creating the team. Adding another button didn't solve the ordering. We put the proposed roster and responsibilities before confirmation.
After hiring, the founder couldn't find the room they liked. We had hidden it in a collapsible panel behind a role summary. We made the office permanent, with Maya as the main contact. An old mobile flex-order rule still put the wrong card first. A phone-width screenshot caught that.
Payments had the same testing mistake. We tested sandbox checkout through a secondary link. The founder used the large price-card button, which still requested an invoice. We repeated the test from the homepage through the main action. Monthly sandbox payment completed. The annual flow reached PayPal address validation and stopped. We recorded zero real revenue from those tests.
- Loved
The useful part of using multiple agents was having someone who hadn't implemented the feature try to break its claim.
For the Buffett-inspired study, a business analyst and a valuation/risk analyst read dated evidence separately, then a manager chose a target. An evidence reviewer found an accounting change that made two cash-flow series incompatible. We rebuilt those inputs. Another reviewer later caught the manager describing a cash-flow bound incorrectly. We preserved the original meeting and displayed the correction.
That study used one retrospectively selected company and four reviews. The team held cash three times and bought at the fourth. It gives us inspectable reasoning and a simulated fill, rather than evidence that we've reproduced an investor's future decisions.
We also rejected a faster manager response that called Wellington's NAV total-return benchmark a price return. Across eight final replies to four fixed questions, median complete-answer time went from 16.05 seconds to 8.51 seconds. The answers were shorter and the network was the same. We kept that scope with the number.
By 7:45 p.m. Pacific on November 18, our audit counted 45 agents across successive groups and 589 confirmed peer sends. We excluded manager traffic and inherited copies. Committee implementation and deployment verification exchanged 33 messages. The lead still had to forward new decisions to existing agents and reconcile files before release. Inheriting a conversation at spawn time didn't create permanently shared memory.
The most useful handoffs had a named owner, an input someone else could repeat and a saved result. The reviewer could say “this returned the wrong decision after I changed funds” and send a reproducible case. That made a disagreement actionable.
- Want
A HarnessRouter preflight on the intended production host would have saved time. We wanted to distinguish an invalid credential, an unavailable harness and a blocked hosting route before a visitor asked Maya a question. The returned response IDs were useful when comparing what the browser showed with what the provider completed.
For our own next pass, custom team creation needs attention first. We'll start at the same public entry point as a new visitor and check several different mandates through the complete roster and confirmation flow. The successful starter roster stays in the demo with its actual scope.
The full journal has the dated execution, payment, voice, deployment and portfolio details, with links to the working product and historical evidence.
Illustrated build journal · Project and demo
For other builders, what do you deliberately test after your agent works locally but before you share the public link? In our case, changing one account state or choosing the more obvious button found bugs that a successful model call had missed.