What you get

  • A level from 0 to 4. It names what your control rests on: nothing on record, written rules, checks, enforcement or measurement.

  • A score from 0 to 100, overall and for each of 16 dimensions. A dimension is a part of the work with agents, from the first written requirement to cost.

  • What to fix first. At most 10 actions in order, safety controls first.

  • Evidence a second person can open. A file, a setting, a hook, a record. A self-rating never reaches a level.

AEI Scan gathers the evidence on your own machines. It opens no source file.

3 kinds of rule

Every rule for an agent is 1 of 3 kinds. The kind decides how far you can rely on the rule.

KindWhat the rule isWhat proves the rule
WrittenA sentence in a file the agent readsThe file
CheckedA mechanism that reads the result after the agent has acted, and refuses the commit, the merge or the releaseThe check, which stops with an error
EnforcedA mechanism that refuses the action before the action happensA hook that can deny the action, a permission rule or a sandbox setting

A hook is a small program the agent's tool runs at a set moment, for example before a command runs. A sandbox is a setting that limits which files and network addresses the agent's commands can reach.

5 levels

Each level rests on stronger evidence than the one below.

LevelNameWhat the level means
0UnrecordedNo evidence was found for this part of the work.
1WrittenThe practice is written down; nothing checks that the agent follows it.
2CheckedA check catches a miss after the agent has acted.
3EnforcedA mechanism stops the action before it happens.
4MeasuredRecords are kept and read, so you can see whether the controls work.

The 16 dimensions

Each dimension gets its own level and score, so you see where control is strong and where it is thin. The last column gives the highest level that dimension can reach.

#DimensionThe question the dimension answersHighest level
1Requirements and recordingIs the work written down before an agent starts, so there is something to check the result against?3
2DecisionsAre choices recorded, and is it written which choices a person makes and which an agent may make?2
3Expected behavior and designIs what a person expects written before the build, and is each expectation tied to a test?2
4Agent briefing and model policyDoes the agent get written instructions, and is the model it runs on fixed by version?3
5BuildDoes each agent work in its own environment, with limits on its commands and on the size of a task?3
6Test and evidenceDoes a check prove the work passes, in place of the agent saying so?4
7Review and mergeDoes a reviewer other than its author (a person, or a model run as a separate reviewer) review every agent change, is it written which changes need a person, and do checks run at each merge?4
8ReleaseDoes a gate stand before the step that reaches your customers, and is its result kept?4
9Safety controlsDo a sandbox, permission rules and hooks stop a harmful action before it happens?4
10Memory and continuityIs what the agent must know in its next session written down, and does each agent record what it learned where the next agent reads it?2
11MeasurementDo you keep and read records of what agents did, what it cost and how quality moved?4
12GovernanceIs it written which AI tools are used on which data and who approves, and can you show which rules are enforced?3
13Token and costIs it written which model does which work, do settings hold the cost of model use down, and are usage records read?4
14Harness depthHow much of the agent tool's own machinery is in use: hooks, connections to outside tools, agents with limited tools?3
15Process guardsHas the team built its own guards, such as 1 guarded script for every delete and time limits on builds?3
16Model conductDoes the team know how each model it uses misbehaves, and does the team act on that?4

The last dimension asks about the model, where the others ask about your process. A model can break a written rule, stop before the work is finished, report work as done when it is not, or give a reason that is not true.

Vendors measure these behaviors in their own models. Anthropic's system cards for Claude Opus 5.5 and for Claude Fable 5.2 and Claude Mythos 5.2 score each model on false claims that a task is complete. A peer-reviewed study of 1,642 runs of multi-agent systems reports 14 kinds of failure, an agent disobeying its task and an agent stopping early among them.

No source we read gives the rate of these behaviors in a team's own work. So the index asks whether you keep a record for each model, whether someone other than the agent checks what it reports, and whether you count.

How a rating is produced

97 checks produce the rating. Each has a level, the evidence it reads and a rule for passing. Evidence reaches a check in 1 of 3 ways.

  • Scan, 35 checks. AEI Scan runs on a developer's machine and reads the files and settings that the agent's vendor documents.

  • Survey, 18 checks. A person on your team answers a question.

  • Interview, 44 checks. An assessor asks to be shown the artifact and grades what is shown.

Evidence outranks description. When your team says a control exists and the scan finds nothing, the check fails, and the report lists it under "Said but not found".

A check with no evidence and no answer is not graded. A dimension with fewer than half of its checks graded is not rated.

The overall level cannot rise above the level of the safety controls, dimension 9, up to level 3. A high score cannot buy a level that the safety controls have not reached. The briefing for the technical team (PDF) gives the arithmetic.

Our own result: AEI 79, Level 1

We ran AEI Scan on the repository of Kyberna, our own product, on 2026-11-16, and 95 of 97 checks were graded. All 16 dimensions were rated, and none came out at level 0.

Read the full report as an example of a report: 81 of the 97 checks come from the process used to build Kyberna, so the result does not validate the index.

What a rating does not tell you

  • No proven link to outcomes. No study read for the index shows that a higher level brings faster delivery or better code.

  • Presence, not effect. The scan shows that a control is present. It does not show that the control works.

  • Judgment. 44 checks rest on a person judging an artifact, and 2 assessors may judge the same artifact differently.

  • One agent, one repository, one day. A rating describes a single coding agent in a single repository on the day of the run. Each coding agent gets its own rating.

  • Your own answers. The survey answers are yours, and nobody checks them unless an assessor asks for the artifact.

Go deeper

Get your own rating

Write to info@prosights.com for AEI Scan, and run it in a repository of your own. The report gives your level, your score and what to fix first.

Ask for AEI Scan