Engineering productivityframework
agentfile — resumable multi-step pipelines with L0–L3 data layers
Problem solved: Ad-hoc multi-step LLM pipelines aren't resumable or debuggable; when step 7 fails you re-run from zero and lose the L0 raw data plus all intermediate state.
What
Generic pipeline framework that structures state across four data layers (L0 raw → L1 cleaned → L2 enriched → L3 synthesis), declares which agents read/write each layer, and persists timestamped workspaces so any step can be re-run independently. Includes an agent-selection matrix and a quality-loop primitive for iterative refinement. Overkill for one-shot single-agent jobs; pays back when the pipeline has 4+ steps or needs to survive crashes.
← All casesC-099