Govern AI agents with controls, not a policy PDF.
Your AI policy covers people. Agents that write to your systems need engineered controls. Build an initial framework, then run it with Hyphen and Redapt.
Identities · Permissions · Owners · Registry · Audit log
Your AI policy covers people. Agents do the rest.
Most teams wrote their AI policy for chat tools. It says what staff can paste and which tools they can use. It says nothing about an agent that updates thousands of records at 2 a.m.
What the AI policy covers
- What staff paste into chat tools
- The list of approved AI tools
- Vendor use of your data for model training
- Who gets a license
What agents do
- Update records in your ERP and CRM
- Change payment terms
- Send email to customers
- Approve refunds
- Call other agents
- Spend tokens on a company card
When agents act, governance is identity, permissions, and logs.
Put these five controls in place before an agent writes to a system of record. Your policy then describes controls that your systems enforce.
- 01Provisioned identities
Agents act as accounts that the company provisions.
Each agent uses its own account. The company provisions it, scopes it, and can revoke it at any time.
- Give each agent its own service identity. Do not use personal API keys or shared admin accounts.
- Start each identity with read-only access. Add scope only when the agent goes to production.
- Keep model and system credentials in one secrets store, so you can rotate or revoke them in one step.
- 02Point-of-action permissions
Read, propose, and write are separate permissions.
A gate checks each action when the agent tries it, not only when the run starts.
- Define three permission levels for each system: read, propose, and write.
- Send each tool call, model call, secret read, and system write through one gate.
- Keep an allowlist for each agent in each environment. Read it at runtime, so a revoked permission stops the next action.
- Reserve a budget for each agent call. Deny calls that go over the budget.
- 03Named owners
Each action type has a named owner.
The workflow shows where the agent decides and where a named person decides.
- List each action type before you build the agent.
- Name one person, not a team, as the approver for each consequential action type.
- Set thresholds, for example a dollar value above which a person must approve.
- Actions that move money always need a person.
- 04One registry
One registry lists each agent.
The registry tells you which agents run, who owns them, and what they can touch.
- Record the purpose, owner, users, systems, permission level, environment, and version of each agent.
- Version prompts, tools, and models together. Record the version for each action.
- Turn off each agent that has no owner. Retire agents that nobody uses.
- 05Append-only audit log
An append-only log records each action.
The log records allowed and denied actions. If the log is down, the agent does not act.
- Write each action, version, rule, and approver to an append-only, tamper-evident log.
- Log denied actions too. They show where the agent differs from the process you expect.
- Block all actions when the audit store is not available.
- Make the log searchable, so you can answer "who changed this record?" in seconds.
Each call goes through one gate.
Your team sets the allowlist for each agent in each environment. The gate reads it at runtime, so a revoked permission stops the next action. If the audit store is down, nothing runs.
Governance gate
Each call goes through it. Nothing goes around it.
On the allowlist
Runs with scoped credentials. The log records it.
All other calls
Stops before it reaches a system. The log records it too.
Append-only, tamper-evident log
Write access comes last.
- 01
Sandbox
Anyone can start an agent with temporary credentials, approved tools, and copies of data. A person reviews each consequential output.
- 02
Shared workflow
When three or more teams want the same agent, it gets one owner, one maintained prompt, evals, and a place in the registry.
- 03
Shadow mode
The agent works the real queue with read and propose access only. A person makes each real change, and the log shows each difference.
- 04
Production
After written sign-off, the agent gets write access for this workflow only. Named people approve each consequential action.
From agent sprawl to one governed workflow.
This is a typical plan for the first function in a large company. Smaller teams move faster. The framework builder sets the schedule for your size.
- Weeks 1–2
Find each agent
Check model provider consoles, service accounts, and AI spend. Include the agents that nobody approved.
- Weeks 2–3
Register each agent
Give each agent one named owner. Turn off each agent that nobody will own.
- Weeks 3–5
Set up three tiers
Write down the access and the data for the sandbox, shared workflow, and production tiers.
- Weeks 4–6
Pick the first workflow
Name each action type in the workflow and one person who approves it.
- Weeks 6–10
Run in shadow mode
The agent proposes and a person decides. The log shows where they disagree.
- Weeks 10–12
Approve write access
Get written sign-off from risk and IT, with names, for this workflow only.
- Each month
Review and improve
Review denials, differences from people, and spend. Retire agents that nobody uses.
The same work, with one agent that anyone can audit
| Before | After | |
|---|---|---|
| Agents for one job | Several copies, each with its own prompt | One agent with one maintained prompt |
| Ownership | Nobody | One owner and named approvers |
| Credentials | Personal API keys, some on personal cards | One company account that you can turn off |
| "Who changed this record?" | Weeks of investigation | One search of the log |
Framework builder
Build your AI compliance framework in three steps.
Choose your industry, company size, and compliance requirements. Get the five controls with clause references, the decisions that stay with people, and a rollout plan for your size.
- 1Industry10 options
- 2Company sizeUnder 100 to 10,000 or more employees
- 3Compliance requirements17 laws, standards, and frameworks, such as HIPAA, SOC 2, EU AI Act, PCI DSS, ISO/IEC 42001
Shared controls, so your teams can build more agents.
Let the people close to the work design agents. Keep the infrastructure that runs them shared. With one platform, you can give teams more freedom in the sandbox.
| Control | Hyphen platform | Redapt team |
|---|---|---|
| Provisioned identities | ENV keeps each agent's credentials end-to-end encrypted, with access logs and key rotation. Roles, SSO, and MFA control who can change them. | Security experts set up a service identity for each agent in your identity provider. |
| Point-of-action permissions | Deploy gives each tier its own environment in your AWS, Azure, or Google Cloud account. Toggle turns an agent's write path on or off at runtime. | Engineers design and build the gate and the allowlist for each agent. |
| Named owners | Hyphen Agent uses the same model: it starts read-only, asks for approval, and checks the plan again before it makes a change. | Fractional CTOs help your leaders map the first workflow and name each approver. |
| One registry | Projects give each agent a home with owners, environments, and deployment history. | Redapt runs the inventory with your team and sets up the registry. |
| Append-only audit log | Audit logs record authentication, administration, access, and changes. Search audit events in the app or through the MCP server. | A 24/7 operations team monitors the systems that keep the gate and the log available. |
AI governance, powered by Redapt.
Hyphen supplies the platform that runs your agents in your own cloud. Redapt helps your team design, build, and operate the controls.
Hyphen
The agent platform
- Deploy runs agents in your own AWS, Azure, or Google Cloud account.
- ENV keeps the credentials of each agent encrypted, logged, and revocable.
- Toggle turns the write access of an agent on or off at runtime.
- Audit logs record access and changes across the platform.
Redapt
The implementation team
- Redapt engineers design, build, and operate the systems that run your agents.
- Your team can bring in Redapt security experts at any time.
- Fractional CTOs help your leaders name owners and approvers.
- A 24/7 operations team monitors those systems.
You keep your cloud accounts, identities, and logs. You can revoke access at any time. Hyphen and Redapt do not certify compliance.
Frequently Asked Questions
What is AI governance for agents?
AI governance for agents is a set of engineered controls. The controls set who an agent acts as, what it can read and change, who approves its actions, and what the log records. A policy document describes these controls. Your systems enforce them.
Why does our AI policy not cover agents?
Most AI policies cover people who use chat tools. They control what staff paste and which tools staff use. A background agent acts on its own in your systems of record, so it needs a control on each action.
What are the five controls?
Agents act as accounts that the company provisions. Read, propose, and write are separate permissions that a gate checks on each action. Each action type has a named owner. One registry lists each agent. An append-only log records allowed and denied actions.
How long does the rollout take?
The plan puts a first workflow into production in 7 to 13 weeks. The time depends on the size of the company. The plan starts with an inventory and ends with write access for one workflow. The framework builder gives you a schedule for your size.
Does the framework builder give legal advice?
No. The builder gives an initial framework and references to related clauses. Your counsel, auditors, and assessors decide what your requirements need.
What do Hyphen and Redapt do?
Hyphen supplies the platform that runs your agents in your own cloud, with environments, encrypted secrets, feature flags, and audit logs. Redapt engineers help your team design, build, and operate the controls. Your team can bring in Redapt security experts at any time.
Do we keep control of our agents and data?
Yes. Your agents run in your cloud account. You provision the identities, you own the logs, and you can revoke access at any time.
Do Hyphen and Redapt certify compliance?
No. Hyphen and Redapt help you build and operate the controls. An independent auditor or assessor decides whether your organization meets a standard.
Hyphen supplies the agent platform. Redapt helps you build and run the controls.
Put agent governance into production.
Tell us where your agents run today. Hyphen and Redapt help you take the first workflow from inventory to production.
- Tell us about your agents and the requirements that apply to you.
- Review your framework with a Hyphen and Redapt team.
- Plan the first workflow, from inventory to shadow mode.