Advisory 2.0: AI Investing Stack That Requests Its Own Tools

When Advisory first launched, the goal was to turn a generic “financial chatbot” into a modular, production-grade AI investing stack - one designed to evolve its capabilities over time.

We set to evolve through a simple principle: when Advisory can’t fulfill a task to increase app performance, it can request a new tool. Over time, this creates a feedback loop where the product continuously expands what it can do - not just in markets and investing, but in product growth and operations too.

For example, Advisory can learn to ask for tools that measure and act on real product outcomes:

• How many users opened a push notification?

• What’s weekly and daily retention?

• Which messages drive engagement?

• Where are new users coming from?

• Should we publish a social post, send a push, or trigger an in-app flow?

• etc.

The 2.0 release takes that idea close to active autonomy: Advisory now runs on both Android and iOS, powered by a team of specialized AI agents that collaborate autonomously in the background to surface exactly the right market moves, news, and portfolio signals for each user.

Advisory 2.0 uses a conversational orchestrator built with Strands Agents that treats other agents as tools. Market data, financial news, user management, and push notifications all run as independent agents behind MCP endpoints, while the top‑level conversational agent decides when and how to call them.

The result: users get proactive, personalized financial insights without having to constantly “pull” information via chat.

Agents-As-Tools: New Multi-Agent Architecture

At the core of Advisory 2.0 is a Strands‑based orchestrator agent that acts as the “brain” of the system. It receives a user’s query (or a scheduled background task) and then routes work to four specialized agents, each exposed as a tool:

• Market Data Agent - provides real‑time prices, exchange rates, technical indicators, and AI‑driven price predictions.

• Financial News Agent - gathers relevant news headlines, company financial reports, earnings results and what analysts are saying.

• User Management Agent - has access to risk profile, goals, portfolio metadata, favourites, and other preferences.

• Push Notifications Agent - composes and triggers Firebase push notifications based on market events and user engagement.

These agents are implemented as standalone services and exposed via a standard HTTP/MCP interface. The orchestrator sees each one as a single “tool” call, even though each agent can internally chain multiple tools, hit external APIs, and perform its own reasoning. This agents‑as‑tools pattern keeps the orchestration layer thin while letting each agent evolve independently.

On a typical complex query such as “Should I scale back my Nvidia position this week in light of my risk profile and recent market events?” the orchestrator will:

1. Ask the User Management Agent for the user’s risk tolerance, goals, and whether NVDA is in favourites.

2. Ask the Financial News Agent for recent Nvidia headlines, analyst moves, and recent events.

3. Ask the Market Data Agent for current price, short‑term technicals (RSI/MACD), and near‑term price projections.

4. Synthesize a short, action‑first answer (e.g. “Trim 10–20% if price stays above X; keep core if Y”) and optionally hand off an event description to the Push Notifications Agent to set up monitoring and alerts.

All of this coordination happens inside the Strands agent loop; the orchestrator simply calls the right agent‑tools and stitches their outputs into a single response.

Self-Aware Tooling: Requesting Its Own Upgrades

Finally, the system is built to evolve. If an agent ever encounters a legitimate user request it can’t fulfill (say someone asks for intraday volatility and no tool provides it), it can invoke a special request_new_tool function. This shared tool lets any agent log a missing capability: it creates a record in the Database with the agent name, the requested capability, an example query, and a justification. The team later reviews these requests to build new tools. In this way, Advisory not only serves users but actively learns where its toolset needs to grow.

Proactive Autonomy: Notifications Without Friction

The biggest shift in Advisory 2.0 is that the agents no longer wait for the user to ask good questions. They collaborate in the background to detect relevant events and then decide to “wake up” the user with a notification.

Behind the scenes, Advisory regularly scans users’ favourite assets and current market conditions. It looks at price moves, market news, volatility, and how unusual the latest activity is compared to normal behaviour. Instead of reacting to every small fluctuation, the system calculates a dynamic significance threshold for each asset and only considers sending a notification when that threshold is clearly crossed. This keeps alerts focused on genuinely meaningful moves while avoiding noise during choppy markets.

Beyond simple alerts, the Push Notifications Agent also fosters community insight through interactive poll votes. When users receive notifications about significant market events, they can vote on questions like "Are you buying?" or "Will this stock recover?" and immediately see how other Advisory users are responding. This gives everyone a real-time pulse on general market sentiment without requiring additional research or guesswork.

State & History: Supabase + S3

To keep everything personalized and context-rich, Advisory stores user data centrally. A Supabase Postgres database holds profile and portfolio metadata: risk levels, goals, favorite tickers, notification limits, timestamps of last alerts, etc. This allows any agent to answer questions like “What’s her risk tolerance?” or “How many shares of AAPL does he own?” without guesswork. For longer-term history and auditability, we use S3 object storage to append a chronological timeline of all past notifications and insights. The chat interface can query or replay this timeline so users see a coherent “log” of advice over time. Together, the database and object store let the orchestrator stitch together answers with full context.

Financial Content with Traceable Sources

When Advisory references market data or news, the content always comes from external financial feeds rather than model guesses. A dedicated MCP Server pulls recent articles and fundamentals, then the agents turn that into short, human‑readable summaries instead of raw link dumps. The UI still exposes underlying sources as clickable cards, but the core experience focuses on concise, trustworthy explanations rather than unverified or hallucinated numbers.

Advisory on iOS and Android

On the client side, Advisory 2.0 now brings the same agent‑driven experience to both Android and iOS. The mobile apps talk to a shared backend for authentication, profiles, and portfolio data, and both platforms plug into the same conversational agent stack and notification pipeline. That means every user, regardless of device, benefits from the same real‑time insights, background monitoring, and proactive alerts.

Because the intelligence lives in the agents and their data layer rather than the UI, the apps stay lightweight: they focus on speed, smooth navigation, and presenting the right insight at the right time.