New: LLM Rankings are live!

See how the top LLMs rank on real infrastructure tasks

Let agents propose infrastructure changes without risking production.

An agent can write Terraform faster than anyone can review it. The danger is a change that reaches production without the policy checks, approvals, and records every other change gets. Massdriver runs agent changes through the same path as everyone else’s.

Where agent changes go wrong

The failures are familiar ones, arriving faster. A plan gets applied without anyone reading it. A database is replaced, and its data lost, because a renamed attribute forced a replacement and the plan said so in one line. A change breaks checkout, and nobody can tell which person or agent made it.

Code review catches less of this than it seems to. The risk depends on what a change does to the infrastructure that is running, and the diff of a Terraform file does not show that. A one-line change to a module input can mean a resize, a restart, or a destroy, depending on the resource and its state.

Agents make each of these more likely because they make more changes, and because an agent rarely hesitates before running the next command.

Safeguards teams put around agent changes

All of these help. Most of them depend on the agent going through the safeguard voluntarily.

Plan output posted to the pull request

What it gives you

A reviewer sees the plan, including replacements, before apply.

What it leaves to you

Someone has to read every plan. An agent with cloud credentials can still apply from its own session.

Lifecycle flags such as prevent_destroy

What it gives you

Terraform refuses to destroy the flagged resources.

What it leaves to you

Protection only where someone remembered to add the flag, in code the agent can also edit.

OPA or Sentinel policy on plans

What it gives you

Rules that block known bad changes at plan time.

What it leaves to you

The rules cover changes that go through that runner. Changes made another way skip them.

Agents only in non-production

What it gives you

Nothing an agent does can reach production.

What it leaves to you

Production changes still happen by hand, so the agent saves you little where it would help most.

Massdriver

What it gives you

Every change, from an agent or a person, goes through policy, approval rules, and the audit log. The agent has no credentials to go around them with.

What it leaves to you

Your team decides which changes need an approver, in which environments.

How every agent change stays on the governed path

01

Policy before every apply

Attribute-based access control decides whether this token may take this action on this instance. Then the bundle’s schema and policy checks (SOC 2, HIPAA, CIS) check the plan before anything runs.

02

A person approves destructive changes

Agents propose provisions and decommissions. A proposal waits until someone with deploy permission on that instance approves it, and stateful resources have deletion protection.

03

Every change is on the record

Each proposal, plan, approval, and deploy is an audit event with an actor. Service accounts are their own actor type, so agent actions are never mixed in with a person’s.

04

Rollback when a change slips through

Deployment history and environment compare show what changed, and rollback returns the instance to its last good state.

A risky change, caught before production

An agent tuning the checkout service changes a database input. In production, that change would replace the database.

  1. 01
    What happens

    The agent proposes the change on the production database instance.

    What the record shows

    The proposal names the agent’s service account and the parameter it changed.

  2. 02
    What happens

    Massdriver validates the input against the bundle’s schema and runs policy checks on the plan.

    What the record shows

    The plan diff and the policy results are attached to the proposal.

  3. 03
    What happens

    The agent has no deploy permission in production, so the proposal waits for an approver.

    What the record shows

    Nothing has run. The database is untouched.

  4. 04
    What happens

    The approver reads the plan diff, sees the replacement, and rejects the proposal.

    What the record shows

    The rejection is recorded on the proposal, and nothing runs.

  5. 05
    What happens

    The agent revises the change so it modifies the database in place, and the approver approves it.

    What the record shows

    The deploy is recorded with the agent as proposer and the person as approver.

  6. 06
    What happens

    If a later change still causes trouble, the team compares environments and rolls the instance back.

    What the record shows

    The rollback is itself a recorded change.

What an approver sees before saying yes

A production approval in Massdriver is a decision about a specific change, with the evidence in one place.

  • The diff. The parameters that changed, the bundle version, and the plan.
  • The policy results. Which checks ran on this plan and what they found.
  • The actor. The person or agent service account that proposed the change.
  • The target. The instance, its environment, and the services that depend on it.

Let agents prove a change before production sees it

Massdriver can fork an environment from production’s blueprint, so an agent can try a change against the same topology, bundle versions, and parameters before anyone reviews it. The approver then sees the plan for production next to the result of the same change in the fork.

How much review a change needs can follow its blast radius. Access policies use blast radius as an attribute, so a change to a stateless preview service can deploy on policy checks alone, while a change to a production database waits for a person.

Questions platform teams ask

Does every agent change need an approval?

No. You decide by environment. Most teams let agents deploy in dev and preview environments and require an approver in production.

Can the agent approve its own change?

Not when separation of duty is on for that environment. The account that proposed a change cannot approve it.

What stops an agent from applying Terraform directly?

It has no cloud credentials. Massdriver holds them and uses them only for deploys that policy allows.

How do we undo a bad change?

Deployment history shows each change to an instance, and rollback returns it to its last good state.

Does this slow agents down?

Only where you put an approver. In environments where agents may deploy, policy checks run and the change applies without waiting on a person.

Do agents get a different process from engineers?

No. An agent change and an engineer change go through the same catalog, policy checks, approvals, and audit log.

When code review is enough

If agents only open pull requests, a person reads every plan, and a person runs every apply, your review process is the safeguard. It stops being enough when agents can act on infrastructure themselves, or when there are more plans than people to read them.