Most agent projects reach the same awkward meeting. IT, risk or the board asks what stops the agent from doing something nobody intended. Up to now the honest answer has often been "the system prompt, and our own testing". That is thin cover once an agent is approving invoices, writing to the CRM or emailing customers without a person checking each step.
In September 2026 NVIDIA and Anthropic gave Claude agents a better answer. NVIDIA announced the Open Agent Safety Platform, an open reference design for limiting what AI agents can do, and Anthropic built Claude Managed Agents directly against it. The Claude announcement describes the pairing in detail. This piece covers what it changes for an Australian business running agents against real systems.
How do Claude Managed Agents and NVIDIA OpenShell work together?
Claude Managed Agents, the API suite for running production agents at scale, now pairs with NVIDIA's open-source OpenShell software. Managed Agents keeps the agent's credentials in a vault, so the agent never sees them. OpenShell sits outside the model and enforces limits on what the agent can execute and reach while it works, even if something in the model's reasoning goes wrong.
The useful idea is the split. One layer controls the keys. A separate layer, which does not depend on the model behaving well, controls the doors. Together they let a business set a hard ceiling on an agent's access, watch what it actually did, and confirm the limits held.
| Question a reviewer asks | Prompt-only agent | Managed Agents plus OpenShell |
|---|---|---|
| Where do credentials live? | Often in config the agent can read | In a vault the agent never sees |
| What limits actions? | Instructions in the prompt | An execution boundary outside the model |
| What if reasoning goes wrong? | Relies on the model noticing | Boundary still enforces limits |
| Can we show what happened? | Depends on your own logging | Watch actions and confirm limits held |
Why infrastructure beats policy documents here
A policy that says "the agent must not touch the payroll ledger" only works if the person building the agent remembered to encode it, and if the model follows it every time. A credential vault plus an execution boundary means the controls exist whether or not the builder remembered. That is the shift: from trusting the model's judgement to stay inside the lines, to a ceiling the model cannot raise.
For Australian organisations this maps neatly onto controls you already report against. Least privilege is an Essential Eight theme. APRA CPS 234 asks regulated entities to show that information assets are protected in proportion to their sensitivity. Privacy Act obligations bite hardest when an agent can read or send personal information. An enforced boundary is something a security review can actually point to; a paragraph in a system prompt is not.
We have written before about why infrastructure, not prompts, separates prototype agents from production ones, and about safety patterns for Claude agents that touch production. This announcement is the vendor-supported version of that argument.
Two questions to ask your build partner this week
If you run Claude agents against anything that writes data, moves money or talks to customers, ask two things:
Where do the agent's credentials live, and can the agent read them? If keys sit in an environment file the agent can open, that is the first fix.
What actually stops the agent doing more than the job needs? If the answer is "the system prompt", Managed Agents plus OpenShell is the upgrade path.
Can you show us a record of what the agent did last week, and evidence that its limits held? If not, observability belongs in the same piece of work.
Related reading on keeping credentials out of agent hands: workload identity federation on the Claude Platform.
What an audit costs compared with an incident
For an agent already in production, working out what it can technically reach is usually a half-day audit, not a rebuild. As an illustrative figure, a focused permissions review for one production agent might run $2,500 to $5,000 for a mid-market firm, with remediation scoped separately. Compare that with the cost of an agent that can write to a finance system with no enforced limit. These numbers are illustrative, not a quote, but the asymmetry is the point.
What not to conclude
An execution boundary does not make an agent's output correct. It limits the damage a wrong decision can do. You still need sensible approval steps for high-stakes actions, testing before go-live and someone who owns the agent once it is running. Nor does this mean every existing agent must be migrated tomorrow. Start with the agents that write, pay or send, since that is where an enforced ceiling earns its keep.
We are folding this into how we scope permissions on new agent builds as at November 2026. If you have an agent in production and are not sure what it can reach, get in touch and we will walk through it, or see our services for how we run agent security reviews.



