Building agent-native biology primitives

TL;DR - Rafflesia is organized around deterministic biology primitives: explicit inputs, typed outputs, provenance, object references, hashes, and warnings that callers can compose.

Biology software has often hidden useful evidence behind dashboards, notebooks, and one-off scripts. That makes it hard for a human, workflow runner, or coding agent to reproduce the same result later. The Rafflesia surface is built around a smaller contract: call one primitive, get one inspectable measurement or artifact back.

Evidence before interpretation

The product boundary matters. A primitive can resolve a target, fetch a sequence, inspect a structure, summarize pLDDT, split regions, search folds, detect pockets, map variants, or export a table. It should not quietly turn those measurements into a biological conclusion.

That separation keeps the system useful in more settings. The engine returns evidence. The caller decides what the evidence means.

Files are part of the interface

Rafflesia commands are designed to compose through JSON, tables, manifests, and object ids. Outputs should include enough information to rerun or audit a result:

• input references and explicit parameters
• database or tool snapshots where available
• content hashes and object references
• typed warnings for skipped or unavailable evidence
• stable dataset names and versions for table exports

The goal is to make each result portable across terminals, scripts, notebooks, and agent loops.

Sharp tools scale better

Agent-native does not require putting an agent inside the product. It means the product exposes narrow, predictable operations that an external agent can call repeatedly without memorizing hidden state.

Small primitives are easier to test, cache, inspect, and replace. They also make failures easier to understand: a missing Foldseek backend should become a structured warning or a clear failure, not a fabricated result.

That is the direction for this blog too: short notes about primitives, data contracts, runtime behavior, and the evidence surface behind Rafflesia.