Enterprise AI platform · 2025 –
Agentic Browser Automation at Scale
Senior AI Engineer, Platform
Challenge
Internal product teams needed browser automation agents that could run reliably at scale: distributed sessions, CDP control, and human-in-the-loop when flows broke. A major revenue line was stuck behind manual workflow gaps.
Approach
- Built and scaled agentic browser automation with Chrome DevTools Protocol (CDP) for deterministic browser control.
- Architected distributed session management for parallel workflows with human-in-the-loop checkpoints for failure recovery.
- Integrated with platform execution patterns so product teams could ship flows, not one-off scripts.
Tech stack
CDP · Distributed browser sessions · Agent runtimes on Kubernetes · Streaming execution · HITL workflow patterns · TypeScript agent SDK
Outcomes
- Unblocked a major product revenue line (eight-figure ARR impact)
- Repeatable browser-agent pattern for product teams, not one-off hero scripts
- Production path from pilot flows to scaled session management
Verifiable patterns (redacted)
Session pool architecture (redacted)
Distributed CDP session manager: browser pods on Kubernetes with session affinity, idle timeout, and max concurrent tabs per workflow. Session lease prevents double-assignment; stale sessions reclaimed via heartbeat watchdog.
HITL checkpoint flow (redacted)
When automation hits an unrecoverable DOM state, workflow pauses and surfaces a human-in-the-loop checkpoint with screenshot, DOM snapshot hash, and suggested recovery actions. Operator resolves or aborts; trace continues with annotated decision for eval replay.
Platform execution hook (redacted)
Product teams register browser flows via agent manifest: CDP steps, retry policy, HITL triggers, and eval golden paths. Execution plane spins ephemeral browser pods with platform auth — no one-off scripts outside the standard runtime.
Context
Internal product organization (customer and product names anonymized).