New Gartner® Report: Secure AI Agents With Identity Governance Foundation

Allow, Needs Approval, or Deny: How NewCore Decides Every Governed Call

Sr Solutions Engineer, NewCoreNov 19, 20265 min read

The decision happens at the call

An AI agent doesn't sign in once and wait. Whether a person prompts each step or the agent works toward a goal on its own, it keeps calling tools until the job is done. Each of those calls is a moment where someone, or something, has to decide: should this run?

Most identity controls aren't built for that moment. Sign-in policies check a person at a login. Access reviews look at grants weeks later. By then, the agent has already acted.

NewCore puts the decision at the call. Every call that goes through the NewCore gateway gets one of three outcomes: allow, needs approval or deny. This post walks through how that decision is made, who makes it and what's left on record. These per-tool decisions are the core of NewCore's Agentic Guardrails. Agent policies are live today.

One policy engine, one place it runs

The decision is made by NewCore's policy engine, and it is enforced at the gateway. The gateway sits between your agents and the MCP servers they use. Agents connect to the gateway, not directly to the MCP servers, so a governed call can't reach a tool without meeting policy first.

By the time policy is checked, NewCore already knows who is calling. Each call arrives tied to a specific agent and the person it answers to, not to a shared API key.

Two properties matter here:

  • Policy sets the boundaries. What's allowed, what needs approval and what's denied comes from rules you write, so you always know where the lines are.
  • The check is inline. It runs before the call reaches the tool, not in a report afterward.

One honest limit: NewCore governs the calls that go through its gateway. Agents that reach tools another way, such as AI built into other SaaS products or a direct OAuth grant left in place, aren't decided by it. Closing those paths is part of deployment design.

Three outcomes, set tool by tool

Every tool an agent can reach is set to one of three outcomes. Anything not set is denied.

OutcomeWhat happens to the callTypical fit
AllowRuns straight through to the toolReads of data the agent's owner can already see
Needs approvalWaits for an authorized person to decideMoving money, deleting records, changing access, sending something outside the company
DenyStopped before it reaches the toolTools that have no business in this agent's hands

Deny by default. An agent can only see tools from connectors whose tool list your security team has reviewed. Tools outside the agent's profile are denied without anyone writing a rule for them. A narrow profile is the cheapest control you have.

Set by consequence, not by vendor. A useful starting point is to ask what the call does if it goes wrong. Reads can usually run. Changes that are easy to undo can run, or need approval for agents that are new or broad. Anything hard to undo should wait for a person.

Who a policy applies to

Policy can be scoped four ways:

  • One agent, for the exceptions.
  • A whole agent platform, such as every Claude Desktop agent, to keep a consistent baseline.
  • A user, when one person's agents need different settings.
  • A group, so a sensitive team can run under tighter settings than everyone else.

When more than one rule matches, only one applies: the most specific. A rule for a named user beats a rule for their group, and a group rule beats one that covers everyone.

Policy also has a ceiling it can't lift. An agent acting for a person is capped at that person's permissions at the provider. If the owner can't write to a system, neither can the agent, whatever policy says and whatever the agent was asked to do.

When a call needs approval

A call set to needs approval doesn't run until an authorized person decides. They see the request in the NewCore App or the User Portal and approve or deny it.

The important part is what happens when nobody answers. The call fails closed: if nobody decides within five minutes, the request expires and the call doesn't run. A request that quietly turns into a yes isn't a control.

Approvals work best when they're rare. If one approver sees a steady stream of requests and approves all of them, that's a signal. The tool should probably be allowed, or the agent's profile narrowed. Save the person for the calls that deserve a decision.

Where intent comes in

A rule knows which tool is being called. It doesn't know why. An approval is where someone checks the why: an inbox assistant that normally reads email now wants to update a deal in your CRM. Did anyone ask it to?

Today, that check is the approver's. Where we're headed, NewCore will weigh it too: what the agent was asked to do, what it has done before and how risky the action is. Routine work goes ahead, and only the exceptions reach a person. Policy still sets the boundaries; intent decides which calls inside them need a human.

Every decision on record

A decision you can't show afterward doesn't help an auditor. Every allow and deny through the gateway is logged, including the automatic ones, against the agent and its owner. Approvals, policy changes and connections sit in the same record.

That record is also a picture of how each agent behaves, and it's how policy gets better. It shows which tools agents actually call, which calls wait for a person, and which agents keep reaching for tools outside their profile. An agent that collects denials is telling you its task, or its instructions, have changed. That's worth a look before it's worth a wider policy.

The same engine can also stop an agent outright. Revocation is checked on every call, so an agent that's been cut off is refused on its next request, not when its token runs out.

Getting started

A practical way to turn policy on:

  1. 1.Start strict on tools that change things. Set them to needs approval and let routine reads run.
  2. 2.Keep profiles narrow. Leave out tools an agent has no reason to use; deny by default does the rest.
  3. 3.Tune from the record. Loosen tools that get approved every time, and look into agents that keep getting denied.
  4. 4.Move approvers to strong sign-in first. An approval is only as strong as the account that gives it.

None of this replaces your IdP. NewCore runs beside the one you already have.

Agents act. People stay accountable. The enterprise can prove it.

Cite this blog

Mengarelli, H. (2026, November 19). Allow, needs approval, or deny: how NewCore decides every governed call. NewCore. https://newcore.com/blog/allow-needs-approval-or-deny-agent-policy

Subscribe to our blog.

New posts, sent when published.