Objects

An Object is a business entity your tenant cares about — Account, Deal, Discovery Call, Ticket. It's the unit of grounded truth an Agent reasons over.

What it does

An Object type declares a business entity once; every document that mentions it can be linked back to a single row, and every Skill that needs context about it gets the same canonical record.

Instances (e.g., "Acme Corp") live as rows in the business_object table, one per tenant, linked to documents in your Sources. An Agent asked "what's happening with Acme?" retrieves the Acme Corp row, not ten emails that all partially describe it.

Folder shape

workspace/<org>/objects/<slug>/
├── type.yaml                   # structured definition (required)
├── classification-prompt.md    # tells the classifier how to recognize this type (optional)
├── evals.yaml                  # classification fixtures (recommended)
└── README.md                   # optional — rationale, edge cases

type.yaml

slug: deal
label: Deal
description: Sales opportunity tracked across CRM, email, and documents
icon: handshake
sourceRelevance:
  hubspot: 2.0
  gmail: 1.5
  zoom: 1.3
  google_drive: 1.2

classification-prompt.md

Optional. Tells the classify_business_object skill how to recognize this kind of object in a document. Free-form markdown — the LLM sees it verbatim.

evals.yaml

Classification fixtures — document in, yes/no out, which instance it should link to. See Evals for the full shape.

Runtime

• classify_business_object skill reads new documents, uses the classification prompt to decide "is this a Deal?", and writes a row if yes.
• generate_summary skill keeps the object's summary field fresh based on linked documents.
• Any Skill can call ctx.retrieve('Acme Corp', { sources: ['hubspot'] }) + resolve matches back to the object row.

Connection to other resources

• Sources — sourceRelevance weights which connectors are most authoritative for this object type
• Skills — typically take an objectId as input and update the object's fields as output
• Workflows — orchestrate classify → summarize → enrich for every new document

Next

• Authoring context — the editor + apply cycle
• Skills — the LLM-powered capabilities that read and mutate Objects
• Evals — classification fixtures