Skip to content

Anthropic Adds Session Budget Caps and Web Domain Restrictions to Managed Agents

· by Pondero Newsdesk

The short version

Three August 2026 additions to Anthropic's Managed Agents platform give operators hard cost controls and network boundary enforcement for production deployments.

Anthropic Adds Session Budget Caps and Web Domain Restrictions to Managed Agents

Anthropic added a hard dollar cap to Managed Agents sessions in August 2026: when a session reaches its budget, it pauses with a budget_reached stop reason rather than continuing to accumulate cost. Two companion changes tightened network and attribution controls for the same platform.

What shipped

Operators can now set a hard dollar cap on any Managed Agents session, priced at public list rates, per Anthropic's release notes. When a session hits its cap, it pauses with a budget_reached stop reason rather than continuing to issue model requests. Operators can change or remove the cap to resume the session. The same budget parameter applies to deployments, where it propagates to every session a deployment starts. This shipped August 7.

On August 18, Anthropic added the ability to lock down which websites the built-in web_search and web_fetch tools are allowed to reach. Setting allowed_domains or blocked_domains on a tool's entry in the agent_toolset_20260401 configs array restricts the agent's external web access per the same release notes. The web_fetch tool also gained a max_content_tokens parameter in the same update, and web_search gained a user_location field.

The third addition, shipped August 11, is an anthropic-workspace-id response header on all Claude API responses. It carries the wrkspc_-prefixed ID of the workspace the request resolved to, including an organization's default workspace.

Why it matters

The budget cap addresses the chief financial risk of agentic workloads: a session that enters a long-running loop and keeps calling the model until someone notices. Before this change, operators had to handle cost-ceiling logic at the application layer or set hard request-count limits through other means. The budget_reached pause-and-resume design does not hard-terminate the session state, so operators can inspect what happened before deciding whether to increase the budget and continue or stop.

The domain restriction feature closes a security gap that matters most in regulated-industry deployments. An agent granted web_search access previously had no built-in limit on which external URLs it could reach. allowed_domains and blocked_domains make that access explicit and auditable at the platform level, shifting the burden off individual application developers and onto a config parameter that compliance teams can review directly.

The workspace-id header is a smaller but complementary addition. It gives compliance teams a per-request workspace attribution signal, which matters when a large organization runs multiple workspaces and needs to tie API usage in logs back to a specific business unit.

What to watch next

Anthropic's release notes describe session-level budget caps but say nothing about per-tool-call limits. Whether that follows is the open question for teams running high-frequency tool-calling workflows where a single call is the cost unit. The domain restriction feature currently uses exact-match lists. Whether Anthropic extends support to wildcard or pattern-based entries would determine how practical the feature is for organizations with large but structured domain sets.

Sources