The Paved Road: compass.next

The gradient an LLM agent actually follows

compassForWorkflow and compassForRun are pure, deterministic functions — no model call, no I/O beyond reading the workflow's own row — that compute a stage, a summary, and a concrete next array of { tool, args, why } suggestions, with real ids already filled into the arguments. LLM agents follow gradients more reliably than they follow upfront doctrine; the compass is that gradient, attached directly to the tool result an agent just received instead of living in a wall of system-prompt instructions it has to remember unprompted.

Loop state as state, not folklore

Every step of the build loop — author, dry-run, debug-run, production run — stamps durable evidence onto the workflow's own settings, keyed to a content hash of the graph, so that evidence goes stale the instant the graph changes underneath it. That's what lets "where am I with this workflow?" be answered by one deterministic read, for any agent or human picking the work back up, instead of being tribal knowledge that lives only in whoever built it last.

Continue