Meet Cody: The First AI Agent to Maintain Agent Access Policy

After a few months in limited availability, Cody is live today for all users of Agen.co, Frontegg's identity-native agentic governance platform built for AI agents rather than people. Cody is an autonomous policy engineer that sees every access request AI agents in an enterprise make, drafts the resulting policy changes, and brings each one to a human manager for approval.

While policy engines for AI agents have existed for some time, a human employee was always responsible for creating and maintaining the policy. Cody is the first AI agent to create and maintain agent access policy based upon what agents are actually requesting to access.

Cody operates from the Agen.co Gateway, which agents' MCP and connector calls to other systems pass through. It views every request an agent makes – not a sampling. Therefore, each policy change proposal arrives with the evidence behind it. This includes a replay of the agent's past calls against the proposed policy. A human administrator approves each change, and a tamper-evident record of each decision is created.

Cody helps CISOs and security teams keep up with governing agents

Cody’s availability meets a growing demand from security teams who are struggling to keep up with governing and securing agent use as it rapidly scales within their enterprise. A recent survey of 300 U.S. CISOs and CSOs for Agen.co’s The State of Autonomous Agent Governance report found that 84% say AI agents in their enterprise’s IT environment take at least 1,000 individual actions a week.

That’s an average of one every ten minutes, around the clock. For one in five, it's more than one a minute. As a result, 57% say keeping agent policies current by hand is getting difficult: 48% report a growing burden, 8% are already falling behind, and 1% say it isn't humanly possible at their scale.

What that looks like in practice

During limited availability, Cody observed a user’s invoicing agent that held four Stripe permissions, including write access. Over 42 days, it had made 1,240 calls. Every one of those was read-only.

Cody proposed removing the write permissions and showed the evidence: replaying the agent's past calls against the proposed policy. Because the change touched payments, two admins had to approve it. The whole review took only a few minutes.

"Agent actions in the enterprise continue to multiply, and the everchanging AI governance policy can’t keep up. Cody does the part nobody has time for. It watches what agents actually request, writes the policy change, and shows you the evidence. The approval stays with a person. That's deliberate: you earn autonomy one rung at a time." - Sagi Rodin, co-founder and CEO, Agen.co by Frontegg.

How Cody works

Cody is hired, not installed. It joins your Agen.co fleet like a worker, with a named human manager. It can read the policy, the usage records, and your identity provider's groups. It can propose changes. It cannot write to the policy.

It learns from four things. The Agen.co Gateway's record of which agent called which tool and how often. An enterprise’s identity provider's groups. The current policy. And every past human decision: including the note a reviewer leaves when rejecting a proposal, which Cody uses and won't reopen without new evidence.

Two ways it updates a policy. Cody learns from two types of triggers. One includes structural changes such as a new identity provider group, a new connector or tool, or a group that was changed or deleted. The other is usage findings from a nightly review: permissions granted but never used, rules with zero traffic, activity no rule covers, rules that conflict, new tools nobody has ruled on, and exceptions approved so often they should become standing policy. Each finding becomes one proposal.

Every proposal is a one-page change. What changes, before and after, why in plain language, the evidence, and how many past calls it would have affected — for example, "0 of 1,240 replayed calls." Cody targets keeping total human review under an hour a week.

Approval is the control. Routine changes need one admin. Sensitive changes — wider access, payments, HR data, secrets, production deletes, exceptions to a guardrail — need two different admins by default. The manager can pause, rename, offboard, or stop Cody at any time, and every action is recorded.

Value on day one. Cody doesn't need weeks of traffic to be useful. On deployment day, with no traffic, it produces a Policy Hygiene Report and a starter policy built from your identity provider's groups and connectors. After about a week of traffic on a connector, it begins tightening that connector against real usage.

An audit trail that can't be rewritten

Every change, reversal, and setting change is written to the Rule Ledger with who decided it, when, and why. Entries are chained to the previous one and signed with a key the product itself cannot use, so no one with database access can rewrite history without it being detectable. One click verifies the ledger.

The ledger exports with mappings to SOC 2, ISO 27001, NIST 800-53, and the EU AI Act. Every change can be reversed, and a reversal is a new entry, never an edit.

Cody is generally available to all Agen.co customers starting today. Agen.co is priced per governed agent, so cost scales with what you actually govern. For a quote and a demo, reach out to the Agen.co team here.