Corey JamesIdeas
Disconnected producers and consumers
Nothing waits on anything else: the method under most of my high-throughput work
Authority outside the model
The model is a nondeterministic worker inside a deterministic system that decides what it may do, what step it is on, and whether it is finished
The copy of record lives off the accelerator
Host memory or the database holds the truth; the GPU or the model holds a bounded working set
The token as the unit
Schedule, budget and account for GPU work in the unit the hardware actually pays in
Tools are the platform
Build the safety machinery once, make it the only path, and let everything else be a declaration
Identity from the caller, credentials on the server
An agent is who it authenticated as, never who it says it is, and it never holds a key it does not need
Compute first, narrate second
Put deterministic evidence in front of the model instead of asking the model for evidence
Measure the instrument
A result is only as good as the proof that the measurement measured the thing
Budgets and refusals
Spend, memory and admission are first-class controls, and the system says no with a number instead of degrading
Reconcile, never resubmit
When a side effect may or may not have happened, write down the uncertainty and check reality before doing anything again
A human at the privileged step
Automate everything up to the one action that carries real authority, and put a person exactly there
Role contracts, with vendors as data
Work is assigned to a role with a contract; which model fills the seat is configuration, and reviewers can be a different vendor from builders on purpose
Separate planes by role
Traffic with different reliability, latency and durability needs gets its own infrastructure, so pressure in one role cannot starve or corrupt another
Append-only, with tombstones and lineage
Durable logs are only ever appended, deletion is a mark, and every derived artefact can prove where it came from
Agents that react to the world
Mailboxes, inbound event sources and standing work orders, so an agent team runs on what happens and never only on prompts