Song Han Team Proposes Coco, a Proactive Desktop Agent with On-Device Context Observation
stanfordnlp · x · 2026-11-16
A paper from Song Han's group at MIT with Stanford's Diyi Yang introduces "deep personalization," a closed loop connecting continuous context observation with continual agent adaptation: the agent observes user activity, selectively retains long-term context, and learns intervention timing from implicit feedback.
Key points:
- The system, Coco, is a proactive desktop agent built around a separate Observer Agent that monitors desktop activity and maintains a compact long-term user modeling context.
- It learns when assistance is welcome from implicit signals like engagement, dismissal, explicit requests, and repeated work patterns.
- Efficient on-device strategies let the Observer Agent run on local AI compute so raw activity data never leaves the device.
- To evaluate at scale without collecting sensitive real-user data, they propose a synthesis pipeline turning existing computer-use trajectories into OSWorld-Persona, a benchmark with personas.
More from coding & agent
- Shreya Shankar and Hamel Husain Release O'Reilly Book 'Evals for AI Engineers' — Dr_Singularity · 2026-11-16
- Developer Runs 42 Parallel Subagents with /swarm in Google's Antigravity — petergyang · 2026-11-16
- Open-source MCP server gives AI agents access to Canada Post: tracking, rates, MyMail — RichardsonDx · 2026-11-16
- Stanford CS329Z 'Engineering AI Agents' course: 12 free lectures from RAG to tool calling — stanfordnlp · 2026-11-16
- Kent C. Dodds: Software Factories Existed 10 Years Ago—Now Agents Work the Floor — thursdai_pod · 2026-11-16
- Kent C. Dodds on AI Engineers: 'software factories existed 10 years ago — human workers' — thursdai_pod · 2026-11-16