Every approach in the guide is scored on the same six criteria, chosen for one question: does it stop an AI coding agent doing something harmful, and can you prove it did? Each criterion gets 0 to 3, for a total out of 18.
Criterion 01
Visibility
Can you see which agents, MCP servers and skills run, on which machines, and what they do?
What it measures
How much of the agent fleet the control can see: which agents are installed, who runs them, at what autonomy level, which MCP servers and extensions they connect to, and the actions they take in a session.
Why it matters
You cannot set policy for agents you do not know about. Developers adopt new agents and MCP servers faster than any approval process, so a control that only sees sanctioned tools leaves the riskiest ones out.
What each score means
0 Sees nothing about agents.
1 Sees an effect of agents (code, processes, credentials) without knowing an agent caused it.
2 Sees agent activity on one channel, such as model traffic, or for one agent.
3 Inventories agents, MCP servers and their actions across the fleet.
Who scores 3
Agent access security brokers (AASB)
Criterion 02
Enforcement
Can it stop or hold a risky action before it runs, not just report it after?
What it measures
Whether the control can act at the moment an agent acts: allow, warn, block or require a person's approval for a command, a file write or a tool call, under a policy the security team sets centrally.
Why it matters
An agent can delete a directory, push to a branch or send a secret in seconds. Detection after the fact is evidence, not prevention.
What each score means
0 Cannot act on agent behaviour.
1 Acts after the fact or on a proxy for the action (a merge, a data pattern).
2 Blocks some actions, or only within one agent's own settings.
3 Allows, warns, blocks or holds for approval at the point of action, from central policy.
Who scores 3
Agent access security brokers (AASB)
Criterion 03
Developer friction
Do developers keep their tools and flow, or do they look for a way around the control?
What it measures
How much the control changes a developer's day: new tools to install, prompts to click through, broken workflows, false positives. Scored so that a higher score means less friction.
Why it matters
A control developers route around protects nothing. Approval fatigue turns every prompt into a reflexive yes, and blocked tools come back on personal machines.
What each score means
0 Breaks common agent workflows.
1 Noticeable: frequent prompts, rerouting or exceptions.
2 Some prompts or setup per developer.
3 Works in the background; developers keep their agents and only see warnings that matter.
Who scores 3
Agent access security brokers (AASB), Endpoint detection and response (EDR)
Criterion 04
Coverage of terminal and MCP actions
Does it understand shell commands and MCP tool calls as agent actions?
What it measures
Whether the control sees and judges the two channels through which coding agents change things: commands run in the terminal, and calls to MCP servers and their tools, whether local (stdio) or remote.
Why it matters
Most agent damage happens through these two channels. A control that only sees model prompts or finished code misses the action itself.
What each score means
0 Does not see terminal commands or MCP calls.
1 Sees their side effects, without the agent's context.
2 Sees them for one agent, or one of the two channels.
3 Sees and judges both, across agents, with the task they belong to.
Who scores 3
Agent access security brokers (AASB)
Criterion 05
Audit evidence
Can you show an auditor what agents did, under which policy, and who approved it?
What it measures
The record the control produces: each agent action, the policy decision on it, who approved an exception, kept long enough and exportable to the SIEM.
Why it matters
Auditors and customers now ask how AI agents are governed. A policy document without a decision log is a claim, not evidence.
What each score means
0 Produces no record of agent activity.
1 Records effects, attributed to the developer rather than the agent.
2 Records agent activity on one channel or for one agent.
3 Records every decision with policy, agent, user and approver, exportable to a SIEM.
Who scores 3
Agent access security brokers (AASB)
Criterion 06
Deployment effort
How much work is it to roll out across every developer machine and keep running?
What it measures
The work to deploy the control to the fleet and keep it current: installs, network changes, per-agent configuration, policy upkeep. Scored so that a higher score means less effort.
Why it matters
A control that takes two quarters to roll out leaves agents ungoverned for those two quarters, and one that needs per-agent tuning falls behind each new agent.
What each score means
0 Major project: new infrastructure and per-team work.
1 Real project: network or per-agent changes on every machine.
2 Moderate: an MDM push or central configuration, plus policy design.
3 Already in place, or switched on with almost no rollout.
Who scores 3
Built-in agent settings, Endpoint detection and response (EDR), AppSec scanning
Scoring rules
Scores are for securing AI coding agents on company-managed developer machines.
An approach is scored on what it does by design, not on what a custom integration could add.
Developer friction and deployment effort are scored so that more points means less friction and less effort.
The six criteria count equally. Approaches with the same total are ordered by coverage of terminal and MCP actions, then enforcement, then visibility.
Each score on an approach's page has its reason beside it, so it can be checked and challenged.
What the guide does not do
It does not score individual products against each other within an approach; listed options are scored on the same criteria as their approach.
It does not invent prices, ratings or statistics. Where a price is not public, it says where to check.
It is reviewed as a whole; the date at the top is the last time every score and fact was checked. See the ranked list and the detailed comparison.