Between August 7 and August 11, the Claude Platform shipped four capability drops for Managed Agents that, taken together, look like a foundation for multi-agent orchestration at enterprise scale. None of them is a headline feature on its own. Together they answer the operational questions that have blocked agent fleets from leaving the sandbox: how to cap spend, how to get a second opinion mid-task, how to version capabilities alongside code, and where the model actually runs.
Session budgets: hard caps on agent spend
Released August 7, session budgets let you set a dollar ceiling on any Managed Agents session. When the cap is reached, the session pauses with a budget_reached stop reason instead of spawning another model call. Adjusting or removing the budget resumes the session. Deployments inherit the same budget and apply it to every session they start.
This is the first time the platform exposes a native financial guardrail at the session level. Previously, teams built their own token accounting or relied on organization-level rate limits — neither of which maps cleanly to "this agent task should not cost more than $5." Budgets work at public list rates, so the math is transparent. For enterprises running hundreds of concurrent agent tasks, this turns an open-ended liability into a predictable line item.
Advisor model: a strategic second opinion mid-turn
Also released August 7, the advisor pattern lets you attach a model at least as capable as the primary agent to a session's multi-agent roster. The primary thread can consult the advisor mid-turn for strategic guidance — effectively a "phone a friend" that doesn't break the prompt cache.
Configure it as a {"type": "advisor"} entry in the agent's roster, naming the model to consult. This is distinct from subagents (which execute tasks) and from the primary model (which drives the conversation). The advisor is a reasoning layer that the primary model can invoke when it hits ambiguity, architectural decisions, or novel failure modes. Early adopters report using Opus 5 as advisor to Sonnet 5 agents on long-horizon refactors where a single wrong architectural choice cascades across dozens of files.
Inference geo: data residency for regulated workloads
The same August 7 drop adds inference_geo control. Set it inside the agent's model object at creation time, or override it per session. Available geographies and their pricing multipliers are documented in the data residency guide (US-only inference carries a 1.1× premium for models released after February 1, 2026).
For financial services, healthcare, and public-sector teams, this removes the last infrastructure blocker: the model now runs where the data is allowed to go. Combined with environment variable credentials in vaults (released June 9), agents can authenticate to internal systems without secrets ever leaving the approved geography.
Skills from GitHub: versioned capabilities alongside code
Rounding out the August 7 release, Managed Agents sessions can now load skills from a mounted GitHub repository. Any skills in the repo's .claude/skills directory are discovered automatically at session start. When the repository updates, the next session picks up the new skills — no agent redeployment required.
This closes the loop between "agent capability" and "software delivery." Skills — folders of instructions, scripts, and resources that extend what an agent can do — were previously uploaded via API or created in the Console. Now they live in the same repo as the codebase the agent works on, versioned with the same git history, reviewed in the same PRs, deployed with the same CI/CD. For teams treating agents as production infrastructure, this is the difference between "a prompt we manage" and "a capability we ship."
Enterprise guardrails: inference hooks and local-session compliance
Two August 5 releases target enterprise governance. Inference hooks (beta for Enterprise) let organizations point Claude at an internal AI security server. Every governed prompt — across claude.ai, Cowork, and Claude Code — is held for an allow/deny verdict before inference proceeds. Requests are signed, failure handling is configurable, and every denial lands in the compliance Activity Feed.
The Compliance API for local sessions (also August 11) extends transcript retrieval to Cowork and Claude Code sessions running on users' machines. GET /v1/compliance/apps/sessions/local lists sessions across the organization; /messages returns the full transcript. Both use the existing Compliance Access Key with the read:compliance_user_data scope.
Together, these give security teams real-time interception and after-the-fact auditability — the pair required for any regulated deployment.
The anthropic-workspace-id header
A smaller but practical August 11 addition: every Claude API response now includes an anthropic-workspace-id header carrying the wrkspc_-prefixed workspace identifier (including the organization's Default Workspace). This lets multi-workspace applications attribute usage, enforce quotas, and route logs without extra bookkeeping.
What this means for agent builders
Four years ago, "agent" meant a prompt with a tool loop. Two years ago, it meant a prompt with subagents and MCP. Today, the platform ships the operational primitives: budgets that pause instead of surprise, advisors that reason without cache invalidation, skills that deploy with git push, and inference geography that satisfies compliance. The pattern is clear — the platform is moving from "run this model" to "run this agent fleet responsibly."
If you're building on Managed Agents today, the August 7–11 window is the inflection point where prototype patterns become production patterns. The advisor pattern alone changes how you architect long-horizon tasks: instead of stuffing all reasoning into one massive context window, you delegate strategic checkpoints to a more capable model that only wakes up when needed. Budgets turn agent spend from a monitoring problem into a configuration parameter. Skills from GitHub mean your agent's capabilities are code-reviewed, not prompt-engineered.
The next gap is observability — the platform shows what the agent did (transcripts, compliance feeds) but not why it chose one path over another. Expect that to close next.
Everything the lab builds in public stays in public. The source is on GitHub, and the current build is a live AI quoting engine for home service contractors.