Context Plumbing (Interconnected) (interconnected.org)

The author describes building an AI system and reframes what’s required for useful AI as “context plumbing”: not just modeling intent (the user’s goal) but continuously moving the right, timely context to wherever an agent runs so it can act instantly and correctly. They argue that winning interfaces are those closest to the origin of intent — always-on devices, embedded agents — and that making agents reliable means supplying background knowledge, tool documentation, user history, tacit assumptions, shared workspace state and session metadata as live inputs, not ad‑hoc lookups. Technically, this shifts architecture away from static CRUD-style apps toward dynamic pipelines that copy and refresh context from disparate sources into the agents’ runtime environment (edge or cloud) to avoid latency, stale data, and repeated expensive queries. Key trade-offs include bandwidth, freshness, compute placement, caching/prewarming, privacy and the incentives around recording context as training data. Practically this implies engineering continuous ETL-like flows, context versioning, and sub-agent orchestration so agents can hill‑climb on intent with the right scaffolding — the “plumbing” that makes Do‑What‑I‑Mean interfaces fast, accurate and competitive.