Most of what an office AI assistant does is small: check the schedule, tidy a folder, summarise meeting notes, fill a form. Send every one of those steps to a cloud LLM and you pay per call, wait on a round trip each time, and watch your company’s data leave the laptop.
Pocket Agent does that work on a small open model (Gemma via Ollama) on my own laptop. Only the hard requests go to Claude. What’s new in this talk? The assistant now drives a real Chrome browser, and it stops for my OK before it submits anything. All live.
Key Takeaways
- Local-First Architecture & Data Privacy: Keep sensitive enterprise data on your laptop by running small open models locally using Ollama instead of leaking context to the cloud.
- Cost & Latency Optimization: Eliminate per-call cloud API costs and network round trips for routine office tasks.
- Hybrid Tiered Routing: Route lightweight desk chores locally and escalate only high-complexity reasoning tasks to frontier cloud models.
- Live Browser Automation with Human-in-the-Loop: Drive a real Chrome browser for end-to-end office workflows with mandatory human confirmation gates before submissions.
Target Audience
- Software engineers, security engineers, AI practitioners, and workplace automation builders interested in local AI, on-device models, and secure agentic workflows.
Prerequisites
- Basic familiarity with LLMs, prompt workflows, and local developer environments (e.g., Ollama or Python).
