Managed Agents got budgets, an advisor, and geo pinning — the guardrails arrived

Managed Agents launched in April 2026 as a way to run Claude as an autonomous worker with a sandbox and built-in tools. The capability was there from the start. The controls were not.

Between August 7 and August 19, 2026, the controls arrived. Six changes, and the theme running through all of them is bounding an autonomous system you are not watching.

Session budgets — the one that matters most

You can now set a hard cap on a session's spend, priced at public list rates.

A session that reaches its budget pauses with the budget_reached stop reason rather than starting new model requests. Changing or removing the budget resumes it. Deployments accept the same budget setting and apply it to every session they start.

This closes the single scariest gap in autonomous agents. Before this, an agent in a retry loop or working on a task harder than you estimated could spend an unbounded amount of money while nobody was looking. The honest answer to "what is the worst case on this agent's bill" was "I do not know."

Now it is a number you set.

Two things to get right:

Pause, not kill. budget_reached is recoverable. The session is not destroyed — raise the budget and it continues. That means a budget can be a checkpoint rather than a guillotine: set it deliberately low on a new agent, let it pause, look at what it did with the money, then decide.

Deployment-level budgets apply per session, not in aggregate. A deployment with a $5 budget running two hundred sessions can spend $1,000. If you need an aggregate ceiling, that is still a monitoring problem, not a budget-setting one. See agent cost control and the Enterprise cost controls that shipped in July.

The advisor, inside a session

A Managed Agents session can now be given an advisor: a model at least as capable as the agent's own, which the session's primary thread can consult mid-turn for strategic guidance.

Configure it as a {"type": "advisor"} entry in the agent's multiagent roster, naming the model to consult.

This is the advisor strategy Anthropic published in April, now available inside the managed harness. The original result: pairing a fast executor with a stronger advisor consulted only on hard decisions improved SWE-bench Multilingual by 2.7 percentage points while cutting per-task cost by 11.9%. Cheaper and better, because most steps in a long task are easy and do not need the expensive model.

If you are running Sonnet 5 agents on long tasks and have not tried an Opus 5 advisor, that is the highest-value experiment available to you this month.

Inference geo pinning

You can now control where model inference runs for a Managed Agents agent. Set inference_geo inside the model object when creating the agent, or override it for a single session.

For anyone who has been told by legal that inference cannot leave a jurisdiction, this is the thing that was blocking you. Pricing varies by geo — check the data residency page before you assume it is free.

Skills from a GitHub repository

When a session mounts a GitHub repository, any Skills in its root .claude/skills directory are discovered automatically at session start and available for that session.

The significance is versioning. Before this, Skills lived somewhere separate from the code they operated on, which meant they drifted. Now the Skill that knows how to work on a repository can live in that repository, reviewed in the same pull request as the code it describes.

If you maintain Skills for more than one project, move them into the repos. It is the difference between a Skill library and a Skill graveyard.

Memory stores in self-hosted sandboxes

Sessions running in a self-hosted sandbox can now attach memory stores. The Python, TypeScript, and Go SDK workers download each attached store into the sandbox at its mount_path and sync the agent's changes back.

Before this, self-hosting meant giving up persistent memory — the two features were mutually exclusive. Organisations that self-host for data-residency reasons were running stateless agents as a consequence. That constraint is gone.

A session viewer you can debug with

The Claude Console session viewer was rebuilt: a timeline minimap, a transcript grouped by model request, and an Inspector panel carrying session details and cost, raw events, per-tool statistics, mounted resources, and per-thread activity.

Per-tool statistics is the line to notice. When an agent breaks, the most common cause is a tool that is failing silently — returning empty rather than erroring, so the model carries on with nothing. A per-tool call and failure count is how you find that in two minutes instead of two hours.

Also: Sonnet 5 pricing stopped going up

Unrelated to Managed Agents but relevant to anyone budgeting against them. On August 10, 2026, Anthropic made Claude Sonnet 5's introductory pricing permanent: $2 / $10 per million tokens. The increase to $3 / $15 scheduled for September 1, 2026 will not happen.

If you built a cost model that assumed a 50% increase in September, take it out. That is a material change to agent economics — Sonnet 5 is the model most agent fleets run on, and a fleet sized against $3/$15 has just gained a third of its headroom back.

What this month adds up to

Managed Agents in April was a capability demonstration. Managed Agents in August is a product an operations person can be accountable for:

Question	Answer as of August 2026
What is the worst case on the bill?	The budget you set
Can I make it smarter without making it expensive?	Advisor on the roster
Where does inference run?	Wherever you pin it
Where do the Skills live?	In the repo, versioned with the code
Can I self-host and keep memory?	Yes, since August 19
Why did it fail?	Per-tool statistics in the Inspector

Every one of those is a question a CFO, a lawyer, or an on-call engineer asks — and until this month, several had no good answer.

Related: Managed Agents · multi-agent orchestration · self-hosted sandboxes · how to evaluate your agents · Anthropic's Managed Agents explainer

Try this today — 30 minutes

Pick the agent you would least like to explain to your CFO — the one whose spend you cannot predict — and put a budget on it.

1. Look up what it actually cost last month. Not your estimate. The number, from the Console.
2. Set a session budget at roughly 1.5× your typical session cost. Deliberately tight. You want it to pause, because a pause is information.
3. Let it run for a week and watch what trips. Sessions that hit budget_reached are telling you something — either the task is harder than you scoped it, or the agent is looping. Both are worth knowing before the invoice arrives.
4. Open the Console session viewer on one paused session and read the per-tool statistics. Look specifically for a tool with a high call count and a low success rate. That is the most common silent cost sink.

The point is not to cap spend. It is to find out where the money goes while the stakes are one session rather than one quarter.

One thing to check while you are in there: deployment budgets apply per session, not in aggregate. A deployment with a $5 budget running 200 sessions can spend $1,000. If you need an aggregate ceiling, that is still a monitoring job.