Skip to content
Spice Rack · the organism
03

the Bloodstream.

the channel

The channel. Every cell acts through one governed tool layer, and a bad call gets refused, on the record.

38%

monitor AI traffic end-to-end

Quoted from published research, not counted from our own systems. That is what the dotted ground marks; our own figures sit on a ruled grid and carry a dated receipt. Attribute as cited.

the unwatched write

Only 38% of organizations monitor AI traffic end-to-end, across prompts, tool calls, and outputs; nearly 20% have no controls at all: no monitoring, no guardrails, no enforcement.

Your AI is acting on your systems, and most teams are not watching it end to end.

Give an agent access and it pokes your databases and APIs directly: it writes a row, calls a service, changes a record. Nothing checks the action on the way through. A bad call just happens, and by the time anyone notices, it has already landed.

where today falls short

A tool call is not a checkpoint.

Today an agent reaches straight into a database or an API and writes. Nothing inspects the call in transit: a malformed write lands and corrupts the shape, an unpermitted write lands because nothing said no, and the ones that should have been stopped leave no record that they were ever tried. Checks that live in each caller are checks that get skipped. They have to live in the path itself.

The circulation

Everything flows through one channel.

Every write the cells make travels one governed channel, checked, gated, and recorded on the way through. Nothing reaches the organism's state unseen. The diagram below shows a write passing the gates.

Blood corpuscles flowing through a branching vessel network (Leeuwenhoek), warm treatment

the answer

One governed channel. Every write is checked in transit, then committed or refused.

Every action a cell takes flows through one governed layer. A write is not a direct poke at the database; it is a typed call that gets shape-checked, then permission-checked, then state-checked against the record itself, before it is allowed to commit. A valid write lands as a durable, audited row. An illegal write is refused, and the refusal comes back typed: it names the exact corrective call to fix it. The bus cannot write bad state, and it tells you how to recover from a rejected write.

One write: shape-checked, permission-checked, state-checked, then committed or refused.

what this teachesEvery action a cell takes is checked on the way through. If it is refused, the error names exactly how to fix it.

the governed action bus🌶️one AI worker (a cell) · one of 351 toolsshaperight format? Zod checkpermissionallowed? policy checkstatelegal now? CHECK · RLS · triggersthen either committed, or refused:the recorda valid write, committed and auditedrefused, and here is the fixRejected: this task cannot run yet.Put it back in the queue first.refusal_code P0004 · requeue_task()doc_ref #closure-recovery

A cell fires a typed call, one of hundreds of governed tools, onto the action bus. The call is shape-checked, then permission-checked, then state-checked against the database rules. An illegal write is refused at the state gate and comes back typed, naming the corrective call to fix it. A valid write passes every gate and is committed as a durable, audited row. Refusal codes and tool names here are illustrative of the mechanism; the real totals are in the record below.

more than a tool call

What a governed bus has that a direct write doesn't.

A direct write is a poke into the void. A governed write goes through one channel, is checked before it lands, and if it is refused it tells you exactly how to fix it.

01

One governed channel.

!Today A lone agent wires straight into each system, and every direct integration is its own unwatched risk.

the cell Every action goes through one bus: a typed tool on one of three governed servers. One place every write has to pass, not a tangle of direct pokes.

351 tools across 3 servers (233 core, 68 infra, 50 client)A real count from our own database, shown as proof rather than a claim. It is a dated snapshot, captured Jul 29, 2026.

one governed channel🌶️a cellbuscoreinfraclientcall
02

Checked in transit.

!Today A lone agent writes first and you find out later, if the bad value is ever noticed at all.

the cell Shape is checked first (a typed schema), then permission (a policy), then state legality against the database itself (checks, row-level security, triggers), all before the write is allowed to commit.

425 CHECK constraints, 84 RLS policies, 138 triggersA real count from our own database, shown as proof rather than a claim. It is a dated snapshot, captured Jul 29, 2026.

checked in transit, then committeda writewriteshapepermissionstatea rowwriteshape, then permission, then state (DB rules)committed
03

A refusal that names its own fix.

!Today A lone agent hits an error and gets a stack trace, and has to guess what to do next.

the cell A rejected write comes back typed: a refusal code, the reason, the corrective call to run, and a doc reference. The bus does not just say no, it names the exact tool to recover.

7 typed refusal-hint classesA real count from our own database, shown as proof rather than a claim. It is a dated snapshot, captured Jul 29, 2026.

a refusal that names its own fixwritestaterefusedrefused, and here is the fixRejected: this task cannotrun yet.Put it back in the queuefirst.P0004 · requeue_task()write

from our own systems

The governed bus, enforcing today

These are not policies in a document. They are the running action bus: hundreds of typed tools across three servers, every write inspected in transit by database checks, row-level securityA rule the database engine enforces itself on every read and write: it filters which rows each person can see, and refuses writes to rows they are not allowed to touch, so the application never has to get the filtering right., and triggers, and a set of typed refusal classes that each name the corrective call to recover a rejected write.

A green value is real, from our own database, as of Jul 29, 2026; in the diagrams, green marks a committed write.

enforcing today

351A real count from the platform's own source tree, shown as proof rather than a claim. It is a dated snapshot, captured Jul 29, 2026.

governed tools that cannot write bad state

  • The core server: the typed record itself. Tasks, decisions, messages and the rest of the coordination store are written through these.
  • The infra server: the machines the organism runs on. Containers, hosts, networks and the game servers.
  • The client server: tools scoped to client work, kept off the two servers above.

a real count from the platform's own source tree, as of Jul 29, 2026

on the action bus

3

governed servers

  • core
  • infra
  • client

at the gate

6

approval policies

changes a human must approve before they run

zoom in

Down to a single refused write

One typed call, checked for shape, then permission, then state, refused at the last mile by the storage engine because its state was illegal, and handed back as a structured error naming the exact tool to fix it. That is the smallest unit of the bloodstream: not a silent failure, a typed refusal that tells you how to recover.

fleet MCP servers → governed tools → hooks and scripts → the github bureau

talk to us

Curious how a layer like this would fit your own records?

Send a note

Refusal codes and tool names in the diagram are illustrative of the mechanism.