What we ran
Timour 01-overview.md · 5 min · last edit Timour Kosters, Nov 18, 14:50For three weeks in Goa, every Edge City India resident who opted in had a personal AI agent that knew what they were looking for, negotiated with other residents' agents, and brought them matches to confirm.
Headline results from the field
Amita 02-results.md · 8 min · last edit Timour Kosters, Nov 18, 14:50Agents were good curators of who should meet. Humans still made the call, and the decisions were often lopsided.
Research design and experiments
Amita 03-experiments.md · 6 min · last edit Timour Kosters, Nov 18, 14:50Live experiments in the village, then a sealed lab where we froze copies of the real agents and swapped the model underneath them.
How we built it
Carter 04-architecture.md · 10 min · last edit Timour Kosters, Nov 18, 14:50One agent per resident, each in its own sandbox, managed by a control plane, matched through Index Network, with every event flowing through one consent-aware door into the research warehouse.
What worked, what broke
Carter 05-worked-broke.md · 7 min · last edit Timour Kosters, Nov 18, 14:50The hosted-agent model held up for three weeks. The pain was in fleet operations, third-party components and making cost visible.
Product map
Timour 06-product.md · 7 min · last edit Timour Kosters, Nov 18, 14:50Everything a resident could touch in V2, what it was for, what we learned, and what we would do with it next.
Next steps and discussion
All 07-next.md · 7 min · last edit Timour Kosters, Nov 18, 14:50What we will send, what we propose to do together, and what we need from Microsoft.