Planning vs. execution
Test where expensive reasoning creates leverage, and where a grounded plan lets cheaper execution perform just as well.
Factory operations · Open research questions
Questions worth
keeping open.
These are working questions, not settled doctrine. Each page is a durable research record—its starting thesis, experiments, evidence, reversals, and eventual decisions—and it changes as the factory learns.
The open research ledger
15 permanent questions
The URLs stay stable while the thinking changes. Open a record to see its current thesis, evidence to collect, status, and most recent revision.
Frontier planning, cheaper execution
Find the boundary between reasoning that needs a frontier model and execution that can safely move to less expensive workers.
Complexity estimation accuracy
Measure whether a grounded plan can predict execution difficulty well enough to improve model routing, budgets, and intervention.
Multiple focused agents
Determine whether specialized roles improve reliability enough to justify their coordination and operational overhead.
Agent communication
Measure when communication transfers essential context and when it becomes expensive coordination churn.
Cost-aware model dispatch
Build dispatch policies that match task complexity to the least expensive model that can reliably complete the work.
Changing prices and batch inference
Turn temporary discounts and batch pricing into deliberate scheduling decisions rather than ad hoc model switching.
Agent decoupling
Find the smallest useful coordination surface between an expanding set of specialized factory roles.
Ephemeral agents
Learn which roles should remain present and which should appear only in response to a specific event or failure.
Introspection and long-term learning
Give the factory enough durable evidence to study its own behavior and improve future plans, dispatches, and environments.
Factory configuration
Balance rigid, observable machinery with flexible agents that can adapt without requiring a new service for every role.
Local and air-gapped inference
Explore the boundary between private, continuously available local inference and elastic cloud capability.
A rich agent playground
Test whether abundant local compute produces better verification and exploration—or merely more activity and noise.
The agent programming environment
Reduce repeated exploration by turning prior searches, failures, and discoveries into context the next agent can retrieve.
The human in the loop
Explore how the human role can move from supervising execution toward product judgment, direction, and creative choice.