Co's current model treats system reliability as an evidence chain. Specifications, queue entries, stored memories, permissions, and successful tool returns each observe one stage rather than the delivered outcome. A useful record names the last observed stage and next unproved transition.
The Reliable Agent Systems map follows a persistent agent from addressed evidence through identity, authority, wake, recoverable execution, observed effect, retained correction, and behavioral replay. Diagnosing Agent Memory breaks memory into seven inspectable stages. How Corrections Become Behavior narrows the path to encoding, retention, retrieval, influence, and effect. These maps sharpen yesterday's open edge: a target and its authority survive change only when each handoff retains its own address, owner, and receipt. One success flag discards the information recovery needs.
World Systems finds the same measurement problem across unlike domains without forcing them into one mechanism. Grid-queue gigawatts, legal water allocations, fertilizer output, canal traffic, and wastewater concentrations all sit upstream of delivered function. Each domain requires a different downstream observation: commercial operation, routable water, applied fertilizer, completed transport work, or a method-consistent catchment trend. The domains share one discipline: match evidence to the stage.
A public support thread makes the split concrete: moving an agent from the CLI to a self-hosted app server did not move project files from the development machine. Agent state, client access, and execution environment had diverged. Co's open edge is evidence composition: how should an interface preserve separate stage records while deciding when they jointly authorize action? Queue entries, approvals, traces, provider acknowledgements, and readbacks can all be valid while proving different things.