PART 03

How It Actually Works

Versioned agents, isolated session sandboxes, permission policies, networking, tools, and events — including the defaults you must change for a production candidate.

Reviewed 31 Aug 2026 · beta · security defaults made explicit

5-Step Lifecycle

A common lifecycle is: create a versioned Agent, create an Environment, start a Session, send and stream Events, then steer or interrupt. Your application must persist the IDs and tested versions it actually used.

01 Create Agent ✓ 02 Create Env ✓ 03 Start Session ✓ 04 Send & Stream ✓ 05 Steer / Stop ✓

Available Tool Categories

These capabilities can be configured; they are not a blanket entitlement. Disable tools a task does not need. The built-in agent toolset defaults to always_allow, while MCP toolsets default to always_ask.

⌨
Bash
Runs commands in the sandbox. Disable it unless required, or require approval and constrain filesystem, packages, network, and credentials.
📄
File Operations
Read, write, edit, glob, and grep within the sandbox's available filesystem. Mount only task-scoped, non-production data.
🔍
Web Search
Server-side search. Configure allowed or blocked domains on the tool itself and treat retrieved content as untrusted input.
VERIFY CURRENT PRICING
🌐
Web Fetch
Server-side URL retrieval. Its domain policy is separate from sandbox networking; allow only sources the task needs.
SEPARATE DOMAIN POLICY
✎
Text Editor
Edits sandbox files. Use synthetic fixtures first, validate outputs, and never treat the sandbox as a cross-tenant shared drive.
🔌
MCP Servers
Connects remote tool providers. MCP defaults to approval, but scopes, server trust, credentials, and returned content still need review.

The Sandbox

Each cloud session receives a fresh isolated Linux container. An Environment configuration may be reused, but sessions do not share that cloud sandbox. Self-hosting moves runtime hardening and isolation duties to you.

Cloud Container
Packages
Base tools plus only the task-specific, version-pinned packages you configure. Review supply-chain risk before adding dependencies.
/workspace/
Session-scoped files and outputs. Use synthetic or minimized data, validate attachments, and delete sessions and uploaded files according to policy.
Claude
The Agent can call only enabled tools. Require human approval for consequential actions; custom tools are executed and authorized by your application, outside Managed Agents permission policies.
🔒 Cloud environments default to unrestricted egress — production guidance is limited networking with explicit allowed hosts

Control Checklist

Platform optimizations such as prompt caching and compaction do not replace explicit application controls.

✓
Tool policy — disable unused tools; use always_ask for risky built-in and MCP actions
✓
Network policy — use limited egress with explicit hosts; configure web-search/fetch domains separately
✓
Resource policy — cap list cost, elapsed time, turns, tool calls, and application request rate
✓
Evidence — capture redacted events, approvals, versions, usage, errors, and final dispositions
✓
Response — support interrupt, kill, credential revocation, quarantining outputs, replay, and rollback

Beta Notice

Currently in Beta
Managed Agents is a beta feature. Include the following header on Managed Agents requests:

anthropic-beta: managed-agents-2026-04-01

Memory-store endpoints currently use agent-memory-2026-07-22 instead. The SDK sets applicable beta headers automatically. The API and pricing may change. Managed Agents is stateful and, per the current overview, is not eligible for Zero Data Retention or HIPAA BAA coverage. Review privacy, retention, residency, contracts, and deletion requirements before using real data.

Sources reviewed 31 Aug 2026: overview, cloud environments, permission policies, and self-hosted security model.