Our Work
Publications, workshops, and milestones from across the initiative.
Featured
Loyal Agents and the Gates Foundation convened researchers, builders and policymakers at Stanford University on October 16, 2026, to work on how a person's agent memory can move with them across services. Speakers included Taylor Kendal (Learning Economy Foundation), Sofie Meyer (Turn.io) and Jiaxin Pei (UT Austin and Stanford University).
MemoryArena tests whether AI agents can use what they learn across sessions to complete later tasks. Agents that score near the top on standard memory tests struggle once memory has to guide real action.
Agentic coding tasks can consume up to 1,000 times more tokens than code reasoning and code chat, and runs on the same task varied by up to 30 times in total tokens. Accuracy peaked at intermediate cost, so spending more did not reliably buy better results.
From the workshop
Archive
Researchers, builders and policymakers met at Stanford to work on portable agent memory.
Agents strong at memory recall often struggle when memory must guide sequential decisions.
An open library of more than 1,000 system prompts from over 400 AI products, audited against a user-centric standard.
Sandy Pentland discussed trustworthy AI agents in trade and finance.
The report extends the Human Context Protocol into education and health applications.
Agentic coding tasks can consume up to 1,000 times more tokens than code chat.
Held at the Gates Foundation on March 11 and 12, 2026, and co-organized with the Gates Foundation, Mozilla Foundation, Stanford HAI, and AWS.
Sandy Pentland spoke on the Imagination in Action panel "Intelligence at Scale: How AI Is Rebuilding Industries and Redefining Human Work," and the team hosted roundtable conversations through the Leaders Beyond Boundaries project.
Identity Management for Agentic AI.
Researchers, policy leaders, and technologists met to discuss trusted agent design.
Governance, trust signals, and potential certification schemes.