Concept: Agent Flywheel

The Agent Flywheel is an engineering and product strategy for artificial intelligence agents. It focuses on building modular, reusable “primitives” rather than bespoke, one-off systems. By designing agent workflows with standardized building blocks, each subsequent agent build becomes faster, cheaper, and easier to implement.

The 9 Primitives (The Agent Skeleton)

A resilient, high-stakes agent architecture is composed of nine key primitives:

  1. Context Pack: Defines the exact data scope and permissions the agent has (what it is allowed to read).
  2. Ingest: Consumes raw inputs and files (e.g., PDFs, emails) and converts them into machine-readable text.
  3. Chunking: Splits long documents into addressable, tagged, and structured sections rather than treating files as huge, unstructured text blobs.
  4. Normalizing: Transforms unstructured strings into clean, typed data (e.g., parsing raw strings into ISO dates, currency amounts, or structured names).
  5. Storing: Storing everything locally (e.g., SQLite, markdown files) so that context persists and nothing depends solely on the model’s transient memory.
  6. Retrieving: Querying structured data by address or similarity search to serve as the exact context for a given step.
  7. Citing: Mapping every claim or drafted text back to the exact source block or document coordinate for validation.
  8. Exporting: Formatting the synthesized findings, draft files, evidence checklists, or timelines into a reviewable “case file” or packet.
  9. Gating (Human-in-the-Loop): A hard boundary where the agent must stop and wait for human review. High-stakes agents are never allowed to sign, pay, submit, or send without human permission.

The Flywheel Effect

When these primitives are built correctly:

  • An email-scheduling agent uses Ingest, Normalize, Gate, and Store.
  • Transitioning to an insurance-appeal agent doesn’t require starting from scratch; it reuses Ingest, Normalize, Store, Retrieve, and Gate, adding only specialized Chunking and Citing logic.
  • A tax-preparation agent then reuses all these assets, making its development highly cost-effective.
  • This modularity allows the use of cheaper, lightweight, or open-source LLMs because the agent’s tasks are constrained to clean, pre-structured data.

References