Claude release notes: Managed Agents load GitHub skills and gain an advisor

Claude Managed Agents can now load a team's own skills directly from a mounted GitHub repository. The same platform update lets the primary agent consult a separate advisor model mid-turn. For teams building recurring agent workflows, the working method can now travel with the code, while difficult decisions get an extra review without moving the entire run to the most capable model.

Claude Managed Agents are Anthropic-hosted agent runs for longer, asynchronous tasks. The platform handles the agent loop, sandbox, tools, and session history, while the team defines the model, instructions, tools, MCP servers, and skills.

Source: Claude Managed Agents overview.

Claude Managed Agents load skills from GitHub

A skill is a reusable set of instructions, resources, and optional scripts for a focused job. Since the August 7 Anthropic Platform release, a Managed Agents session can discover skills at the root of a mounted repository under .claude/skills/<skill-name>/SKILL.md.

This makes one practical change: a review routine, release procedure, or document workflow can be versioned beside the rest of the work. The agent sees each skill's name, description, and path when the session starts, then reads SKILL.md when the task matches. There is no separate upload step or skill ID.

Discovery runs once at session start and uses the branch or commit checked out for that session. A change pushed during a run therefore appears in the next session, not the current one. Discovery also requires the agent's read tool and works for GitHub resources in cloud sandboxes, not self-hosted sandboxes.

Source: Agent Skills in Claude Managed Agents.

The repository also becomes part of the agent's trust boundary. In practice, review .claude/skills, mount a verified commit, and give the GitHub token only the permissions the task needs. The agent can then use real operating procedures without anyone pasting secrets into chat.

Source: Accessing GitHub for Claude Managed Agents.

An advisor gives the primary agent a second opinion

An advisor is a separate model that the session's primary thread can consult mid-turn for planning, getting unstuck, or reviewing work before delivery. It is configured as an advisor entry in the multiagent roster and must be at least as capable as the agent's own model.

This is not another general worker. Only the primary thread can consult the advisor, and the advisor cannot receive normal agent messages or delegate work. A failed consultation does not fail the primary agent's turn. Its usage does count toward the session's total cost.

The useful pattern is a bounded escalation: let the regular model follow the team's skill and do the work, then ask for advice at an architecture choice, a contradiction in the sources, or a final quality check. That gives the workflow a stronger review step without running every part on the most capable model.

Source: Multiagent orchestration and advisors in Claude Managed Agents.

More controls in the August 7 release notes

Anthropic also added hard session budgets and a choice of model-inference geography. A session budget is measured at public list rates and pauses new model requests with budget_reached when the cap is reached. Requests already in flight finish, so the final cost can land slightly above the cap. In a multiagent session, the primary agent, subagents, and advisor share one budget.

The inference_geo choice can be set in the agent's model configuration or as a session override. It is separate from where the sandbox runs. Together, these additions produce a clearer operating model: versioned working methods, a named advice step, a spending cap, and an explicit choice for model-inference geography.

Source: Claude Platform release notes, August 7, 2026.

Source: Session budgets in Claude Managed Agents.

Try this prompt this week

Human step: Mount a trusted repository with a reviewed .claude/skills/<skill-name>/SKILL.md, select the branch or commit to test, and configure an advisor only when the workflow has a real escalation need. If you use a session budget, set it when creating the session.

Read the task brief, repository policy, and available .claude/skills.
Name the matching skill and report its path and commit.
Identify at most one decision that merits advisor review, or explain why none is needed.
Propose a run plan with tools, approvals, and a clear stop condition.
State the budget assumption and evidence to retain in the run receipt.
Change nothing yet. Return the plan for human review.

Good output should:

• Cite a real skill and the version loaded by the session.
• Use the advisor for a named difficult decision rather than every step.
• Separate the agent's tool permissions, human approvals, and spending cap.
• Return a short run receipt with sources, assumptions, and a stop condition.

Tool Forge fits teams that want to move a working routine out of an individual prompt and into a versioned agent workflow that can be reviewed after every run.