이미지: Claude 릴리스노트
Summary
- Claude Managed Agents sessions can now have a hard budget cap set. Once the budget is exhausted, the session pauses in a budget_reached state without issuing new requests.
- An advisor model feature has been added, letting the agent's main thread consult a more capable model for strategic advice mid-execution.
- Control over the inference execution region and automatic skill loading from GitHub repositories were also launched alongside these features.
Session budget caps: a hard cap to prevent cost overruns
Claude Managed Agents sessions can now have a budget cap set. The configured amount acts as a hard cap based on public list rates, and once a session reaches this limit, it pauses with a budget_reached stop reason without initiating new model requests. Changing or removing the budget automatically resumes the session.
Deployments also accept the same budget settings, with the same cap applied to every session that deployment starts. Details are available under the Session budgets section of the official documentation. This is a feature that can be directly used to prevent unintended cost spikes in long-running agent workflows.
Advisor models: requesting strategic advice mid-execution
An advisor feature has been added that lets a session's main thread consult a separate model for strategic advice mid-execution. The model designated as advisor must be equal to or more capable than the agent's own model. To configure it, add an {"type": "advisor"} entry to the agent's multi-agent roster and specify the name of the model to consult.
This feature allows a more capable model's judgment to be inserted at complex decision-making or planning stages that a single model may struggle to handle. The advisor tool can also limit the model's output length via the max_tokens parameter, which is used to reduce response latency and output token costs. The Give the session an advisor section of the official documentation provides specific configuration instructions.
Inference region control and automatic GitHub skill loading
This update also includes a feature for directly controlling the region where model inference runs. Users can set inference_geo inside the model object when creating an agent, or override it for a single session. Available regions and their respective rates can be found under the Data residency section of the official documentation. This feature plays an important role in enterprise environments with data sovereignty requirements.
Automatic skill loading from GitHub repositories was also launched. When a session mounts a repository, skills located in the .claude/skills directory at the root of that repository are automatically discovered at session start and become immediately usable by the agent. This allows skill sets to be managed purely through repository structure, without separate manual registration.
Context from previous updates: rapid expansion of the agent platform
Even before the August 7 update, Claude Managed Agents and the platform as a whole saw a series of changes. On August 5, Inference hooks launched in beta for Claude Enterprise organizations, and the Claude Opus 4.1 model (claude-opus-4-1-20250805) was deprecated the same day, with all requests to it now returning errors. Anthropic recommends upgrading to Claude Opus 5, and researchers can apply for continued access through the External Researcher Access Program.
In July, Claude Opus 5 (claude-opus-5) launched. It supports a 1M-token context window and a maximum output of 128k tokens, priced at $5 per million input tokens and $25 per million output tokens — the same as Claude Opus 4.8. Additionally, Claude Sonnet 5 (claude-sonnet-5) carries introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which it will shift to $3 per million input tokens and $15 per million output tokens. Both models support a 1M-token context window and a maximum output of 128k tokens.
