Concept reference
Browse the concepts that make up a domain-specific agent harness. For a working example with definition files, follow the implementation guide.
Concepts
Section titled “Concepts”Core definitions
PersonaA versioned definition of a domain, its teams, and their reusable work.Domain ontologyThe entities, fields, and relationships that give a team’s data a shared meaning.TeamsThe owners of goals, agent placements, handoffs, and shared resources.AgentsThe workers that execute playbooks inside a team.GoalsTeam-scoped outcomes that select and invoke the playbooks needed for the work.Playbooks (skills)Reusable workflows with declared inputs, outputs, execution rules, and the skills an agent needs to carry them out.Model providersServices and adapters that connect the harness to a model through an authorized account.ModelsThe language or multimodal models that the workspace binds to required capabilities.PromptsInstructions and selected context that the harness assembles for a model call.ToolsThe callable capabilities a harness exposes to an agent within a playbook’s scope.Playbook delegationHow a goal or playbook assigns bounded work and receives a result it can use.Portable playbooksThe versioned format and execution contract for moving a playbook between compatible harnesses.
Execution and governance
Governed executionHow a human decision, a sandbox rehearsal, and a live action fit into a playbook run.RunsAn attempt to execute a goal or playbook with pinned definitions and workspace evidence.Human reviewThe domain judgment and exact approvals that let a team govern a run.Sandbox rehearsalAn isolated test of a proposed action that records evidence without production effects.Execution receiptsEvidence from a live operation that helps a runtime verify and reconcile the result.EvalsVersioned cases that verify a persona’s expected behavior before release.Context and conversationsThe shared knowledge, messages, and files that give a team the background for a run.ApprovalsAuthorization from the assigned person to perform a specific protected action.Lifecycle hooksChecks the harness runs at declared points before or after playbook work.Code executionHow the harness runs scripts in isolated compute with explicit inputs and permissions.Scheduled workHow a team starts work on a schedule or resumes a playbook after a wait.Coach computer useHow Gabriel Coach guides people through a shared browser tab and hands control back for review.
Data and integrations
ConnectorsDeclared external operations and the workspace connections that authorize them.VaultWorkspace storage that keeps credentials behind authorized connections.Records and listsList definitions describe fields while workspace records hold the live values.PipelinesStages and transitions that describe how a record moves through a process.AppsBuilt-in views and custom interfaces over authorized team data and actions.Data feedsVersioned sources that read, validate, and map observations into the domain.SignalsConditions that turn observations into reasons to propose work.Files and artifactsHow authored files, working files, and shared outputs are stored and accessed.
Infrastructure and distribution
RuntimeThe software that imports definitions, binds resources, and coordinates execution.Agent harnessThe system that runs the agent loop, manages context, and coordinates model and tool calls.WorkspacesInstallations that hold local accounts, permissions, records, and run history.Resource binding and keysHow portable definitions identify resources and how a workspace supplies the models, accounts, people, and destinations needed to run them.Local AIConnect the harness to a model on a Mac mini, DGX Spark, or another private inference host.RegistryThe lock that pins linked definition files to reviewed content and Git revisions.Packaging and installationHow authors version definitions and adopt them in a workspace.Live sync and presenceThe events that refresh open views and show who is working in a shared workspace.MarketplaceDiscover and review reusable playbook, team, and persona releases.
How the concepts work together
Section titled “How the concepts work together”- Personas and workspaces traces a definition from app inventory to an installed workspace.
- Teams, goals, and playbooks shows how a team selects and reuses work.
- Connectors and approvals follows a declared operation through account binding and review.
- Records, pipelines, and apps traces a record from the source to a team view.
- Runs and collaboration covers the shared context and activity a team sees while work runs.
- Signals, evidence, and evals shows how observations and expert feedback inform later work.