Notes on building agent infrastructure.
Founder notes from building Lobu in the open — memory, isolation, automations, and the infrastructure underneath company agents.
The Modern Data Stack for Agents
Connectors are CDC, entities are the modelled tables, Automations are dbt models on a schedule. Most of the analogy transfers. It breaks in four places: the loop has to close, derived state is only ever a cache, every read carries who it is for, and you cannot re-run a transform that sends email.
READ ▸
Self-Improving Agents Are Event-Sourced
Intelligence stopped being the blocker. Context is. A filesystem gives an agent no schema and no audit trail, so Lobu puts an append-only log in the middle: memory writes are commits, Automations are mutations, and approval is the merge.
READ ▸
The Context Layer Is an Event Store
Tasman says no product can build your context layer because the content doesn't exist. Half right. The governed why is a versioned, access-controlled event store that mirrors the warehouse instead of replacing it — and I built the Kelder scenario to find where that holds and where it doesn't.
READ ▸ The Modern Data Stack for Agents
Self-Improving Agents Are Event-Sourced
The Context Layer Is an Event Store
The Agent Loop Is the New SaaS
Shopify's Aquifer, in the Open
Filesystem vs Database for Agent Memory
MCP Is Overengineered, Skills Are Too Primitive
