Engineering leadership

How do I make product requirements machine-readable for AI agents?

Structure the facts an agent must act on, while keeping the human explanation. Use typed artifacts, explicit links, stable identifiers, lifecycle state, and acceptance criteria that can map to evidence.

Model decisions, not document shapes

A long Markdown template is still unstructured if an agent must infer which sentence is authoritative. Give each requirement a type, owner, state, version, parent, demand links, and bounded acceptance criteria.

Preserve prose for rationale and nuance. Structure the fields that control selection, authority, traceability, and verification.

  • Identity

    Assign a stable identifier that survives title edits and tool synchronization.

  • Relationships

    Represent parent, persona, problem, dependency, and code links as typed edges.

  • Verification

    Write acceptance criteria as observable outcomes and classify exact contracts separately.

Expose one read and write contract

Agents need a documented interface for search, context reads, work packets, and proposed changes. Do not make each client scrape a different project-management view.

Helen’s Foundry exposes its artifact graph through MCP. Agents can retrieve current context and propose bounded changes while the workspace retains human authority over material premises.

Validate the graph

Machine-readable data can still be incomplete or contradictory. Run structural and semantic checks so agents stop on bad context instead of executing it.

Related reading

Give every agent the current product picture.

Join the Helen’s Foundry early-access program.

Join the waitlist