01
Separate the stream from the state
The event log is append-only and authoritative. Canvases, artifacts, pins and tasks are
projections that can be rebuilt. A product that treats the current view as the truth cannot
replay, cannot audit and cannot recover.
02
Parity in the stream, asymmetry in the power
Agents are addressable senders with audit records, and they can be assigned work exactly
like a person. They never authenticate and never hold human permissions. Every agent has a
human owner who answers for it.
03
A pin is a placement, never a grant
Visibility is access. If placement granted access, any member of any team could publish
into any other team by pinning. Data is executed as the viewer, so the same pin can
legitimately show different rows to different people.
04
Never let the model be the security boundary
An agent that can write SQL can read anything the connection can reach. Datasets are
validated specs, tools are admitted by capability, and grants are checked where the write
happens — not where the prompt is built.
05
Fail closed, and make the failure visible
An isolation floor never falls back to a weaker placement. A broken refresh never blanks a
pinned report. A missing permission renders a permission state, not a partial answer. An
admin cannot believe they have a boundary they do not have.
06
Separate the axes, then compose them
Engine and placement. Capability and data scope. Canvas, artifact and pin. Runtimes and
worker processes. Collapsing an axis buys a week and costs a migration.
07
Split only on a concrete trigger
No service is separated in anticipation. One process owns the schema, the auth plane and
the event log until a real scaling, runtime or compliance trigger says otherwise — and the
seam is already in place for when it does.
08
Keep the vocabulary honest
A tenant is an Organization, a Team is a Team, and Workspace means only the UI shell. A
word that means two things is a bug that ships in every future conversation about the
system.