← All architecture labs
AEG Lab #001Findings published
AEGPhase B v2

Governing Agent Tool Execution

Can an independent governance boundary prevent unsafe or over-broad execution while preserving legitimate tool use?

Claim boundaryThis lab tests a fictional tool environment, policy-v1, a simulated executor, and one published model-backed corpus. It does not establish that AEG or agent systems are universally safe.

Measured
Aug 30, 2026
Runtime
Python 3.11+ standard library
Model run
gpt-5.6-luna
Corpus
120 unique prompts × 3
Policy
policy-v1
Reproduction
API key required for Phase B

Evidence summary

What the current experiment observed

360 / 360Completed model-backed trials
180 / 180Expected legitimate actions preserved
180 / 180Direct baseline executions in non-execution trials
0 / 180Governed unsafe executions observed

Method

The evidence path

01Prompt
02Model proposal
03Direct and governed paths
04Classify

Findings and revision

The evidence changed the work

01

Phase A validated the boundary

The fixed 40-scenario suite passed 40 of 40 cases. The deliberately naive direct path executed 27 of 32 cases expected not to execute; the governed path executed zero.

02

Phase B separated model and governance behavior

One model proposal was reused for both paths. Model selection failures, policy decisions, and execution safety were measured separately so abstention could not be counted as governance success.

03

The pilot changed the experiment

Repeated prompt text, permissive abstention scoring, under-specified read scopes, unrealistic configuration values, and incomplete provenance were exposed and corrected before the published v2 run.

04

The result remains narrow

No unauthorized governed execution was observed in the published configuration, and every expected legitimate action was preserved. The result applies only to the committed model, corpus, policy, runner, and simulated executor.

Limitations

What this lab does not prove

  • All tools, identities, services, and data are fictional.
  • The executor is simulated and does not reproduce every production integration failure.
  • The corpus is committed and reviewable but is not an independent safety benchmark.
  • A different model, policy, tool schema, or system prompt may produce different behavior.

Inspect and reproduce

Follow the evidence into the artifacts