Two Agents, One Truth
People will run several AI agents. Fabric gives them one record of what the person wants and what each agent did.
The problem. Each vendor's agent keeps its own memory. When a person has a travel agent and a home agent, neither knows what the other was told or did, so they act on conflicting wishes and miss each other's changes.
What this demo shows. Two separate Claude Code agents shared one Fabric Recall server. The home agent, asked to book a ride timed to a flight it did not book, searched Fabric, found the travel agent's booking and the airline's words, and recorded that its ride rests on them. It also found that Sam had told the travel agent 'aisle' and itself 'window', and asked Sam to choose. When the airline moved the flight, Fabric recalled the travel agent's flight and the home agent's ride (and the hotel the travel agent tied to the flight): none missed, none wrong.
How it works:
- One record, many agents. Every wish and action names the agent that recorded it; any agent can find and build on the others' work.
- Clashes surface, not resolve. Wishes told to different agents are compared; the person decides.
- The cascade crosses agents. A change reaches every action resting on it, whichever agent took it.
The honest part. The person, accounts and changes are staged; both agents are Claude Code, told apart by name; they ran one after the other on one machine. In the first two-agent run the staged trip was dated in the past and the flight change was defined too narrowly; both were fixed before this run, and both runs are on record.
Fabric is the shared record a person's agents keep together: what was asked, who did what, and what to redo when it changes.
Two separate AI agents ran one person's errands through one Fabric Recall record: one found and built on the other's booking, caught the person's conflicting wishes across the two, and when the flight moved, Fabric recalled the actions of both.
A synthetic person and pretend accounts, made up for this demo: no real bookings, no money. The two agents (separate Claude Code processes, ) and everything they did are real, run once.
1. What Sam told each agent
Each agent heard only its own errands and wishes. Both recorded them in Sam's words, in the one record they share. Highlighted: two wishes about the same thing, told to different agents.
2. One agent builds on the other's work
The home agent was asked to book a ride timed to a flight it did not book. It searched Fabric, found the travel agent's booking and the airline's own words, and recorded that its ride rests on them. Each step below names the agent that recorded it.
3. The world changes
Press a change. The retraction cascade follows it through both agents' records.
4. What the home agent told Sam
What this does not show
- The person, the accounts and the two changes are staged; the changes were applied after the run by a fixed key.
- Both agents are Claude Code, told apart by name; agents from different vendors would connect the same way, and that is not shown here.
- This is the second two-agent run. In the first, the trip was dated in the past (a fault in the staged world: the home agent rightly refused to book a ride for a past date) and the flight change was defined too narrowly to reach what the travel agent had relied on; both were fixed before this run, and both runs are on record.
- The agents ran one after the other, sharing one server on one computer; a hosted record with sign-in is the next phase.
- The hotel night was recalled with the flight because the travel agent tied it to the flight; the key counts it as neither a must nor a must-not.