Define the recurring process, objective and failure cost.
Documentation hub · current architecture reference
Build, connect and govern agents with the same vocabulary the runtime uses.
This hub explains the core Createting primitives and the path from configuration to controlled execution. Detailed reference pages are still expanding; the terminology below follows the current platform architecture rather than a marketing abstraction.
Architecture-awareNo invented API surfaceStatus-sensitive
01 · Quickstart
Start with a bounded workflow, not an abstract autonomous agent.
A useful first deployment has a clear trigger, known systems, measurable outcome, explicit exception path and an authority boundary the agent must not cross.
Tools, skills, knowledge and model routes are scoped to the agent.
Configure autonomy, conditions and approval boundaries before external actions.
Review traces, failures, approvals and evidence before widening scope.
02 · Core primitives
Separate who acts, how work is structured and how it is operated.
Agents are the actors: they interpret tasks and use configured models, tools, skills and knowledge. Agent Studio sets their autonomy and permissions; a Thinkway can structure the method they use.
ActorDurable objectives with tasks, workers, dependencies, approvals, state transitions and operator intervention.
Durable workPathways are rule-based workflows: connect predefined steps, conditions, tools and approval gates using the familiar workflow-automation model.
DeterministicThinkways are reusable reasoning workflows inside agents, not separate agents. Nodes describe work stages with instructions, tools, skills and conditions. The agent applies the method; autonomy is configured separately.
Reasoning assetParent coordination for intake, decomposition, bounded worker selection, progress and synthesis.
CoordinatorOperator surface for durable work, traces, environments, approvals and intervention.
Operations03 · Execution surfaces
Different interfaces can share one governed execution model.
Resources remain scoped to the agent and can carry conditions, restrictions and approval requirements.
Sessions are modeled as runtime capabilities with state, environment traces and operator handoff boundaries.
Voice can gather context, invoke permitted resources, escalate and hand off without bypassing action authority.
Worker routing and agent-to-agent primitives retain explicit identity, lifecycle and responsibility boundaries.
04 · Governance
Authority is evaluated at the action boundary.
The shared policy gate can evaluate tool, knowledge, web search, memory write, browser, computer, external API, subagent and artifact actions before a decision is made.
The configured capability and current policy permit the prepared action without an approval pause.
ProceedExecution pauses at a human approval boundary; the prepared action can be bound to an approval fingerprint.
ReviewThe action is not permitted in the current scope or policy state. Explicit deny conditions remain hard boundaries.
Block05 · Current deep dives
Use focused pages when the mechanism matters more than the overview.
GovernancePermissions, approvals and action authorityOpen →OrchestrationDurable multi-agent coordinationOpen →BrowserObserve, prepare, review, execute and verifyOpen →VoiceConversation connected to governed actionsOpen →ArchitectureWorkflow Architecture ClassificationOpen →
Approval binding
Action Authority & Approval Binding Protocol
Open →
Orchestration
Agent Orchestration Architecture
Open →
Voice evaluation
Voice Agent Production Evaluation Standard
Open →
Resolution
Resolution Contract Specification
Open →
Memory
AI Agent Memory vs RAG
Open →
Platform evaluation
Enterprise AI Agent Platform Evaluation
Open →
Documentation boundary
See what is available before you architect around it.
Agent Builder / native capabilitiesImplemented
Tenant isolation / RLS auditImplemented
Browser runtime foundationImplemented
Environment operations sliceIn progress
Governed runtime decision bindingDEV_READY · hardening open