PROTOTYPE · HUMAN-IN-THE-LOOP AGENTS
Automation that knows when to ask permission.
HomeOps Agent explores explicit boundaries for useful household automation.
Problem
Household automation needs to distinguish a safe suggestion from an action that requires a person’s approval.
Engineering question
How can an agent use tools while keeping permission, failure, and audit behavior explicit?
What I built
A prototype with explicit agent tools, approval gates, deterministic scenarios, failure handling, audit history, a visible web dashboard, and a Strands adapter.
Architecture
The deterministic Good Night workflow and Strands-facing action tool share a host-enforced permission boundary. Only three exact low-risk device/action pairs can reach the simulated mutator. Lock and access requests stop for human review; no approval-resolution flow is implemented.
Evaluation
The reviewed October 18, 2026 implementation CI run passed 37 deterministic tests on both Python 3.11 and 3.12. Chromium scenario-parity coverage checks that the browser reflects the same permission and failure behavior.
Guardrails and boundaries
Current device integrations are simulated for reproducible testing. Agent-supplied approval claims cannot authorize a protected action. A separate host enforcement log records decisions; the agent-authored audit narrative is not authorization evidence.
Failure evidence
Failure handling is modeled as a first-class path; this prototype does not establish reliability in a real household.
Tradeoffs
Simulation makes behavior repeatable but does not validate real device behavior or operational resilience.
Current limitations
The device layer is simulated. There is no production deployment, real-device integration, approval-resolution flow, or demonstrated live Strands model/tool-calling quality. The host and its policy code are trusted; this is not an operating-system sandbox.
What comes next
Extend validation before connecting any real device or durable action.