Goals

A goal describes an outcome for one or more AI agents to pursue. It provides a shared definition of success while leaving execution details to the agents responsible for the work.

Goals should be specific enough to evaluate, durable across individual agent runs, and explicit about any constraints that affect completion.

Each goal belongs to an organization and records a short title, a long-form detailed description that may contain Markdown, an optional accountable organization member, an ordered list of titled success criteria, and four orthogonal dimensions:

• Lifecycle status is active, paused, completed, or archived.
• Health is on_track, at_risk, or off_track, or absent when health has not been reported. Completing or archiving a goal clears current health.
• Timeframe is either continuous or deadline-based with a target date. Existing records that predate this model retain an explicit unspecified timeframe until edited.
• Current evaluation is met, not_met, or unknown as of a specific timestamp, or absent when the goal has not been evaluated.

Drafting is a client-side editing state; a new goal must choose a continuous or deadline-based timeframe and enters the durable domain in active status when it is published. Organization-defined labels provide stable taxonomy and can be renamed without changing goal references.

Status is a projection of an append-only update sequence. Creating a goal records revision 1; every subsequent status report supplies a summary, a long-form Markdown-capable description, the expected current revision, and an idempotency key. A report may also include a new evaluation and its independent asOf timestamp. Appending the report, advancing lifecycle status and revision, and materializing supplied health and evaluation values happen atomically. A report without new health or evaluation retains the corresponding current value. Health is recorded in update history; terminal statuses clear the goal's current health. Metadata updates cannot change status, health, or evaluation. Goals can be archived to preserve their history or permanently deleted with their complete update history when that record is no longer needed.

Each update records the authorizing user separately from the immediate actor. The actor can be a user, client, or verified agent, with an optional agent run identifier. Authentication provenance records the verified session, API-token record, or OAuth client that established authority. MCP client name and version are retained only as descriptive client metadata; self-reported values do not establish identity.