Experiments
Experiments
One question, one attempt, one honest account of how it went. Each write-up carries its hypothesis, its architecture, what came out of it, and what it does not prove.
Reading what the customer actually sent
Customers do not send clean text. They send a scanned price list, a photographed sign and a PDF exported from a word processor in 2011.
multimodaldocumentsragknowledgeMCP as the tool boundary for a multi-tenant agent platform
MCP standardises how tools are described and transported. It does not standardise who is allowed to call them - and in a multi-tenant product that is most of the work.
mcpagent-infrastructuretool-useintegrationsA browser agent for portals that have no API
Reading from portals was tractable. Writing to them was not - and the honest conclusion was to stop trying to make it autonomous.
computer-usebrowser-agentsautomationsafetyOne agent, four inboxes: sharing context across channels
The channel adapters were the easy part. Deciding that the caller and the web visitor are the same person is the actual research problem.
omnichannelagent-memoryidentityconversational-aiBarge-in without false interruptions
Interruption handling is a two-sided error problem: stop too eagerly and the agent is skittish, stop too late and it talks over the caller.
voice-aiturn-detectionbarge-inWhere the milliseconds go in a phone call
A turn-level latency budget for a telephone voice agent, stage by stage - and the three places where overlapping work buys more than a faster model does.
voice-airealtimelatencyarchitectureBooking an appointment end to end, without a human in the loop
The model part is trivial. The hard parts are concurrency, timezones, and what the agent says while a calendar API is thinking.
ai-agentstool-usevoice-aiintegrations
An experiment is not a product announcement. Where a number appears as a diagram it is a design budget unless a sentence says otherwise, and where we do not have a figure we are willing to defend, we say that instead of publishing one.