Claude Managed Agents update memory, domain controls, and session viewer

Self-hosted sandboxes and memory

Work done in a self-hosted sandbox — an isolated execution environment that developers run on their own infrastructure — can now be saved to memory. This is done by attaching a resource of type memory_store when creating a session.

session = client.beta.sessions.create( agent=agent.id, environment_id=env.id, resources=[{ "type": "memory_store", ... }] )

Because the memory store is specified at session-creation time, teams running agents on their own servers rather than in the cloud can now carry that output forward into subsequent sessions.

Domain controls: limiting what agents can see

The web_search and web_fetch tools have gained allowed_domains and blocked_domains parameters, letting developers directly specify the scope an agent can search and retrieve from.

tools=[{ "type": "agent_toolset_20260401", "configs": [{ "name": "web_search", "blocked_domains": ["competitor.example", "content-farm.example"], }, { "name": "web_fetch", "allowed_domains": ["platform.claude.com", "internal.wiki"], }], }]

The published example blocks competitor domains and content farms via blocked_domains, while allowing only an internal wiki and Anthropic's official documentation via allowed_domains. For work such as market research, this means developers can control where an agent pulls information from with just a few lines of code.

Console session viewer redesign

The console screen for monitoring multi-agent sessions, where several agents run simultaneously, has also been rebuilt. It now includes a minimap assigning one lane per agent, a streaming transcript grouped by each agent's iterations, and an inspector showing dollar costs broken down by thread and by session. This addresses a pain point where, as the number of agents grew, it became hard to see at a glance which thread was driving costs — solved here by restructuring the screen itself.

Update	What changed
Memory	Self-hosted sandbox output can be saved as a memory_store at session creation
Domain controls	web_search and web_fetch now support allowed_domains/blocked_domains
Session viewer	Per-agent lane minimap + iteration-grouped transcript + dollar-cost inspector

How to use it

All three features are applied at the API code level, so the entry point is the development environment.

1. To use memory integration, add a resource of type memory_store to the resources array when calling client.beta.sessions.create. Subsequent work in the same sandbox will reference this memory.
2. For domain controls, when creating an agent, add blocked_domains or allowed_domains to the web_search and web_fetch entries within the agent_toolset_20260401 configuration in the tools array. Both parameters can be used at the same time.
3. The session viewer requires no separate setup — the new screen appears as soon as you access the console. Per-agent lanes appear in the minimap on the left, and clicking a specific segment brings up the transcript and cost for that point in the inspector.

For example, a team running multiple market-research agents simultaneously could add competitor sites to a blocklist while allowing only an internal wiki, then use the session viewer to immediately see which thread each agent spent the most on. Developers prototyping in a self-hosted environment can now carry the output of a previous session forward via memory and continue using it in the next run.

Editor's take

Two updates landed ten days apart — budget caps and an advisor model on August 9, then memory, domain controls, and the session viewer on August 19 — signaling that Anthropic is pushing Managed Agents toward being production infrastructure rather than an experimental API. The budget cap controlled "how much" an agent could spend; domain controls now govern "where" it can look. Read together, the direction is clear: rather than simply granting agents autonomy, Anthropic is progressively narrowing the boundaries within which that autonomy can operate.

From experience putting agent harnesses into production, the biggest early cost leak tends not to be the model itself but web calls where an agent repeatedly scrapes unnecessary pages. A single blocked_domains setting to block competitor sites or content farms may look trivial, but in practice it simultaneously improves search result quality and cuts wasted tokens. The session viewer's per-thread dollar-cost display serves the same purpose — what matters isn't how many agents are attached, but which thread is driving costs, and that's the information needed to decide where to set a budget cap.

For teams in Korea running multi-agent pipelines, domain controls are the place to start with this update. Check the logs now for what agents are actually pulling from the web, and if certain domains keep showing up, adding them to a blocklist first will translate into real cost savings. Memory integration is immediately useful only to teams already operating self-hosted environments, while the session viewer redesign pays off most noticeably in setups running three or more agents simultaneously.

Given that Anthropic has now filled in budget, domain, and visibility in sequence, the next update is likely to address controls over inter-agent communication or collaboration rules. As multi-agent sessions grow more common, "who can instruct whom" is set to become as much a management concern as "who can see what."