Beyond the API: Building Truly Autonomous Agents with Workers

The standard mental model for AI applications is still deeply rooted in the "Client-Server" era.

User types prompt -> Server runs LLM -> Response returns.

This is fine for a chatbot. It is insufficient for an Agent.

If an "agent" only works when you explicitly tell it to do something, it's not an agent—it's a tool. A true agent works while you sleep. It anticipates needs, monitors streams, and acts without a trigger from the user.

At Elani, we realized that building a "Personal AI Productivity Agent" meant throwing out the standard API-first architecture. We didn't need a better chatbot; we needed a fleet of autonomous workers.

The Problem with "Chat" Architecture

In a typical chat app, the user is the Driver. The system is dormant until an event (a message) occurs.

But productivity isn't reactive.

• Research needs to happen before the meeting, not during it.
• Organization needs to happen continuously, as emails flood in.
• Briefings need to be ready when you wake up, not generated on demand while you wait.

To achieve this, we moved the "Driver" from the user to the infrastructure itself.

The Worker Fleet

Elani is built on a monorepo of specialized Cloudflare Workers. Unlike a monolithic API that waits for HTTP requests, these workers operate primarily on Cron Triggers and Queues.

1. The Heartbeat (scheduler-worker)

The scheduler-worker is the pacemaker of our system. It doesn't wait for you. It wakes up on a schedule:

• Daily Scan: At 4 AM, it wakes up to process your calendar and inbox for the day ahead.
• Item Research: Every hour, it checks the queue for items that need deep research (e.g., "Who is John Doe from the 2 PM meeting?").
• Reorganization: Periodically, it looks at your scattered tasks and topics and asks, "Does this structure still make sense?"

2. The Brain (shared-utils/classifiers)

When a worker wakes up, it doesn't just "ask an LLM." It invokes specific, structured classifiers.

We don't send free-form text to the model. We use strictly typed Zod Schemas to force the AI to think in data structures.

• TopicNotificationClassifier: Determines if a topic update is worth your attention right now.
• OnboardingFactsClassifier: Extracts durable facts about you (timezone, role, preferences) to update your long-term memory.

3. The Memory (D1 & Vectorize)

Context is maintained across workers using D1 (our SQL database) and Vectorize (our embedding database). When scheduler-worker runs a scan, it doesn't just look at the last hour. It pulls from the User Graph—understanding that "The Project" refers to the Q3 initiative you mentioned three weeks ago.

Why "Silent" Intelligence Matters

This architecture allows for what we call Silent Intelligence.

When you log into Elani, you aren't greeting a blank chat box. You are stepping into a workspace that has already been prepared for you.

• The research is done.
• The emails are drafted.
• The calendar conflicts are flagged.

You verify. You approve. You execute.

Conclusion

We believe the future of AI isn't a smarter chatbot that you talk to for hours. It's a system that requires less talking.

By shifting from "On-Demand" to "Always-On" worker architecture, we are building software that respects the most finite resource you have: your attention.