Browser AI has always had a fundamental limitation: it only works when you are actively using it. You open a tab, ask a question, get an answer, and close the tab. The AI goes dormant. If you wanted to monitor a price drop on a laptop, track new job postings matching your criteria, or watch for competitor product changes, you had to remember to come back and ask again. Repeatedly. Manually. That era is over.

With Tensor 2.1, we are introducing Persistent Background Agents — always-on AI workers that operate in the background, check conditions on a schedule you define, and take action when those conditions are met. They survive tab closures, browser restarts, and even system reboots. They are the first truly autonomous agents built into a browser extension.

How Persistent Agents Work

At the core of persistent agents is a three-phase lifecycle we call the Wake-Check-Act loop. Understanding this loop is key to understanding what makes persistent agents different from simple browser automations or cron jobs.

Phase 1: Wake. Tensor uses the Chrome Alarms API to schedule periodic wake-ups. Unlike setTimeout or setInterval, Chrome Alarms persist across service worker restarts. When you create a persistent agent, Tensor registers an alarm with Chrome at your specified interval — every 5 minutes, every hour, once a day, whatever suits your use case. When the alarm fires, the agent wakes up.

Phase 2: Check. Once awake, the agent evaluates its condition. This is where the AI comes in. Rather than checking a simple boolean flag or comparing two numbers, the agent can perform sophisticated evaluations: navigating to a page, reading its contents, extracting structured data, and determining whether the condition you described in natural language has been met. For example, if your condition is "the price of the RTX 5090 drops below $1,500," the agent will navigate to the product page, locate the current price, parse it, and compare it against your threshold.

Phase 3: Act. If the condition evaluates to true, the agent executes its action. Actions can range from simple notifications to complex multi-step workflows. The agent might send you a desktop notification, compose and send an email, fill out a form, add an item to your cart, or trigger an entire orchestration pipeline.

Creating Your First Persistent Agent

One of our core design principles is that creating agents should feel like having a conversation, not writing code. Here is how you would create a price monitoring agent in Tensor:

You: "Monitor the RTX 5090 on Best Buy every 30 minutes.
      When the price drops below $1,500, notify me
      and add it to my cart."

Tensor: "I'll create a persistent agent for that.
        Here's what I'll set up:

        Schedule: Every 30 minutes
        Target:   bestbuy.com/rtx-5090
        Condition: Price < $1,500
        Actions:  1. Desktop notification
                  2. Add to cart

        Should I activate this agent?"

You: "Yes, activate it."

Tensor: "Agent 'RTX 5090 Price Watch' is now active.
        I'll check every 30 minutes and notify you
        when the price drops. You can manage this
        agent in Settings > Agents."

That is the entire setup process. No configuration files, no dashboard to navigate, no API tokens to manage. You describe what you want in plain English, and Tensor translates it into a persistent, scheduled agent.

Under the Hood: Chrome Alarms and Service Workers

Building persistent agents on top of a Chrome extension presented unique engineering challenges. Chrome extensions run in a service worker context that Chrome can terminate at any time to conserve resources. This means you cannot rely on long-running processes or in-memory state.

Our solution uses three key mechanisms:

The result is an agent that feels always-on even though, at the system level, it wakes up only when needed, performs its check, and goes back to sleep. This design keeps resource usage minimal — a persistent agent checking every 30 minutes uses less than 2 MB of memory on average.

Real-World Use Cases

During our beta, users created agents for a wide range of tasks. Here are the categories that emerged most frequently:

Price Monitoring. The most popular use case by far. Users set up agents to watch prices on Amazon, Best Buy, Newegg, airline booking sites, and hotel aggregators. One beta tester saved over $400 on a flight to Tokyo by setting an agent to watch Google Flights hourly for two weeks.

Job Scanning. Developers and designers created agents that monitor job boards — LinkedIn, Indeed, Greenhouse, and Lever — for postings matching specific criteria. The agent can filter by title, location, salary range, and even company culture keywords extracted from the job description.

Competitor Tracking. Marketing teams used persistent agents to monitor competitor websites for pricing changes, new product launches, blog posts, and press releases. One user set up an agent that checks five competitor homepages daily and produces a summary email every Monday morning.

Inventory and Restock Alerts. Sneaker collectors, GPU hunters, and limited-edition enthusiasts created agents that watch for items to come back in stock. When the agent detects availability, it can add to cart and send a notification within seconds.

Content Monitoring. Researchers and journalists set up agents to watch government databases, regulatory filing pages, and news feeds for mentions of specific topics, companies, or keywords.

Agent Management and Observability

Persistent agents run autonomously, but you remain in control. The Agent Manager panel in Tensor settings gives you full visibility into every active agent. You can see when each agent last ran, what it found, what actions it took, and what its next scheduled run is.

Each agent maintains a detailed execution log. If an agent fails — because a page structure changed, a site is down, or a CAPTCHA appeared — the log records the failure reason and the agent retries on the next scheduled run. After three consecutive failures, the agent pauses itself and sends you a notification asking you to review it.

You can also set agents to "quiet hours" — time windows during which they still check conditions but suppress notifications. This is useful for overnight monitoring where you want the agent to accumulate findings and present them all in a morning summary.

Privacy and Resource Considerations

Persistent agents run entirely on your machine. No data about your agents, their conditions, or their findings is ever sent to our servers. The AI processing happens through your own API keys with the provider of your choice. This means your monitoring activity is completely private.

Resource usage was a primary design concern. Each agent wake-check cycle typically takes 2 to 5 seconds and uses minimal CPU and memory. We batch agent checks when multiple agents are scheduled close together, and we stagger wake times to avoid thundering herd problems. Even with 20 active agents, Tensor's background resource footprint stays below 50 MB of memory.

What Is Next

Persistent agents in Tensor 2.1 are just the beginning. We are already working on agent collaboration — the ability for one agent to trigger another, creating chains of autonomous behavior. Imagine a price-monitoring agent that, upon detecting a deal, triggers a research agent to compare reviews, which then triggers a purchase agent to complete the checkout. Fully autonomous shopping, grounded in your preferences and constraints.

We are also exploring cross-device agent sync, where agents created on your desktop can be monitored and managed from your phone. And we are building a community agent library where users can share their most useful agent configurations as templates.

Persistent Background Agents represent a fundamental shift in what a browser extension can be. Your browser is no longer just a tool you use — it is a tool that works for you, even when you are not there.