Control outside the agent
The proposal places enforcement in external control components. The agent’s own instructions are not a sufficient security boundary.
Read slide 22026-09-03Technical material · AI Alliancen
Agent Harness · reference architecture
A vendor-neutral proposal for a shared control framework with a signed manifest, isolated execution, central gateways and reproducible evidence.
From the material
Three things
to take with you.
An explicit manifest
The contract describes permitted tools, data and operating limits, providing a common basis for control and review.
Read slide 8Evidence belongs in the architecture
Traceability, testing and reproducible documentation inform decisions about whether to allow greater autonomy.
Read slide 17Bring it into practice
A place to begin.
- Draw the control layer separately from the agent and model.
- Describe permissions, data access and stop conditions for one concrete case.
- Define the evidence required before the next maturity decision.
Still open
Questions to explore.
- Which rules need technical enforcement and which require an accountable person?
- How are denied actions and failures tested?
- What evidence would a municipality need before taking over operations?
Read the source material
The presentation.
