Skip to main content
Agent Studio executes deterministic multi-agent state graphs with tool calling, human-in-the-loop approvals, and query-time permission verification. Tasks that exceed a single prompt-response turn are decomposed into graph nodes executed across specialist agents.

​
Graph Primitives

State Graphs

Define execution DAGs with branching logic, error recovery pathways, and typed state transitions.

Tool Calling

Equip agents with verified read-only MCP tools or custom API integrations that require signed session authorization.

Human Approval Gates

Pause stateful workflows before executing sensitive actions until an authorized operator approves the turn.

​
Agent Security Boundaries

  1. Non-Escalating Scopes: Agents execute under the permissions of the calling principal. An agent cannot call endpoints or access collections forbidden to the initiating user or service key.
  2. Deterministic Step Limits: Workflows enforce hard recursion and turn budgets (default: 15 turns) to prevent infinite reasoning loops.
  3. Audited Tool Payloads: Every parameter passed to external tools and every response returned is persisted to the immutable audit log.
Last modified on September 27, 2026