Taskmosis

Taskmosis | AI Browser Automation (Chrome Extension)

Taskmosis automates real web workflows directly in the browser, letting an AI agent complete tasks end-to-end while you watch every action. It's positioned around security and transparency: the agent runs in your browser session and you maintain visible control in real time.

Key Capabilities

• Cross-Platform Integration: Runs on top of the apps people already use (Google, Microsoft, Amazon, Instagram, Uber, Shopify, Airbnb)
• Data Extraction: Extract company names, descriptions, and contact details from webpages and export to Google Sheets
• Smart Form Filling: Auto-fill complex forms using saved profile data (work history, education)
• Intelligent Navigation: Navigate across pages to find specific info like pricing details or documentation
• Workflow Automation: Mass job applying, data collection, product stock monitoring with notifications

Differentiators

• Runs in Your Browser: Operates within your authenticated session with no credential handoff risk
• Visible Control: Watch actions live with full transparency and control
• Cross-Platform Expansion: Desktop and cloud versions in production (macOS, Windows, cloud access planned)

---

Agentic AI Architecture

Agentic Execution Model

Taskmosis is built around an agent loop that can plan, act, observe, and recover across multi-step browser workflows:

1. Plan: Breaks a goal into smaller actions (navigate, search, extract, fill, submit)
2. Act: Executes steps directly in the browser
3. Observe: Reads the page state after each action to confirm progress
4. Recover: Detects mismatches (wrong page, missing element, validation errors) and adapts the plan instead of failing silently

This is what makes it "agentic" - it's not just generating text, it's driving a sequence of decisions and UI interactions until the objective is completed.

Accessibility Tree Parsing

Instead of relying only on fragile CSS selectors, Taskmosis uses the Accessibility Tree to understand page structure:

• Extracts semantic roles (button, textbox, link, heading, menu item)
• Uses accessible names and labels (aria-label, associated label text)
• Builds a stable map of interactive controls even when class names change
• Result: Higher resilience on modern web apps where DOM structure and styling classes change frequently

Context and Memory Management

To complete long workflows without losing track, Taskmosis maintains structured context at multiple levels:

• Task Memory: User's goal, required fields, constraints, and completion criteria
• Step Memory: What has already been tried, what succeeded, and what failed
• Page Memory: Current URL, detected page type, key elements found, and state signals
• User Data Memory: Saved profile fields for form filling (work history, education, contact info)

This enables:

• Continuing after page redirects or login steps
• Filling multi-page forms without repeating fields
• Explaining what it did and why if something fails

Chrome DevTools Protocol Integration

Taskmosis uses Chrome's automation interfaces (CDP) to execute actions with precision:

• Action Execution: Click, type, select, scroll, focus, navigation control
• State Verification: Confirm element visibility, enabled state, and post-action changes
• Network-Aware Logic: Detect page loads, XHR completions, route changes in SPAs
• Observability Hooks: Log timing, failures, retries, and per-step results

Safety, Control, and Observability

• User in Control: Operates transparently with visible actions and checkpoints
• Guardrails: Confirmation prompts for destructive actions (submits, purchases, deletes)
• Telemetry: Captures metrics like step duration, failure reasons, and recovery attempts

Implementation Details

• Uses a structured "world model" of the current page derived from the accessibility tree plus DOM hints
• Uses a planner that selects tools (navigate, search, extract, fill, submit) and validates each step with assertions
• Uses short-term memory for the active run and optional persistent memory for user-provided profile fields
• Uses retries with backoff and alternative strategies when element targeting fails