Agents, Environments & Sessions
Managed Agent Architecture
Building production agents requires managing environments, tools, state, and execution loops. Anthropic's agent architecture patterns provide a standardized blueprint for deploying persistent agents.
Core Agent Primitives
| Primitive | Definition | Role |
|---|---|---|
| Model & Instructions | System prompt + reasoning model | Strategic decision maker |
| Tool Suite | JSON Schemas for functions, APIs, and sandboxes | Execution capabilities |
| Session Memory | Persistent conversation state and workspace context | Contextual continuity |
September 2026 Update: The 'auto' Permission Policy
In the September 10, 2026 update to Claude Managed Agents, Anthropic introduced the auto permission setting. Previously, developers configured tools with either static allow or synchronous ask human-confirmation gates. With auto, the Anthropic execution harness evaluates each outgoing tool or MCP call against contextual threat profiles, parameter bounds, and historical drift in real time. Low-risk read-only commands proceed automatically, while write/destructive actions pause the session and generate a verified approval request.
SOE & AOE Operational Runbook for Claude Agents
For enterprise-grade Standard Operating Environments (SOE) and AI Operations Engineering (AOE), production deployments should enforce:
- SOE Baseline: Standardize runtime environment variables (
ANTHROPIC_API_KEY,ANTHROPIC_BASE_URL), pin SDKs (@anthropic-ai/sdk >= 0.39.0), enforce Zero Data Retention (ZDR) contracts on corporate workspaces, and scope outbound network egress via private proxies. - AOE Tool Gatekeeping: Deploy the
autopolicy across all third-party MCP connectors, pair with real-time audit logging via OpenTelemetry, enforce exponential backoff with full jitter on 429 rate-limit responses, and set deterministic max_turn limits to prevent runaway iteration loops.