03 / AI & WORKFLOW INTELLIGENCE

AI Agents Grounded in Deterministic Execution

Moving beyond simple chat prompts and wrapper interfaces. My focus in AI engineering is autonomous agent architecture: combining LLM reasoning with deterministic browser runtimes, semantic blueprints, and runtime self-healing systems that adapt when user interfaces change.

TECHNICAL INTEGRITY: All capabilities described below reflect verified system implementations. No exaggerated benchmark claims, simulated intelligence, or unsupported autonomous capabilities.

FEATURED AI AGENT SYSTEM · WORKFLOW REPAIR & HEALING

IntentIQ · Autonomous Web Agent

“Record Once. Train Never. Test Anywhere.”

WATCH DEMO↗IN-DEPTH ARCHITECTURE STUDY→

0:00 / 0:00IntentIQ Autonomous Execution Engine

Near-Viewport Autoplay

Live Agent Execution: IntentIQ interpreting semantic blueprints, reasoning over DOM mutations, and executing multi-step workflows without brittle XPath reliance.

01 · REASONING & DISCOVERY

Blueprint Engine Memory

Humans don't think in CSS selectors; humans think in workflows. Instead of hardcoded DOM paths, IntentIQ stores high-level semantic intent, making recorded test actions resilient across framework updates.

02 · RUNTIME SELF-HEALING

Judge & Repair Engine

When an element moves or dynamic classes mutate, the Judge Engine detects failure, invokes OpenAI reasoning, and inspects visual screenshots to repair broken paths in real time.

03 · UNEXPECTED ENGINEERING

AppleScript OS Dialog Handlers

Browser drivers cannot dismiss OS-native popups (Save Password, Location, Notifications). IntentIQ implemented background AppleScript daemons to continuously clear operating system modals.

OpenAI GPT-4oPlaywrightBrowserUseSemantic BlueprintsAppleScript DaemonsComputer Vision Verification

ARCHITECTURAL LESSONS

What Building an AI Agent Taught Me About Software Survival

Running Tests vs. Verifying Truth

Triggering actions via Playwright was straightforward. The hard engineering question was: “How does the agent know whether it actually succeeded?” An HTTP 200 response does not mean a form submitted correctly. That forced the creation of screenshot diffing and multi-agent validation.

Determinism vs. Cost Trade-offs

Relying purely on LLM vision calls per DOM click is prohibitively slow and expensive. IntentIQ executes deterministically via Playwright whenever possible and only delegates to LLM reasoning when a path breaks — balancing sub-second speed with autonomous adaptability.

← Previous Pillar: Workflow Automation Next Pillar: Cloud & DevOps→