Source: Matthew Purcell, "I gave an AI the keys to my smart home", Unprompted newsletter on LinkedIn, 2026-11-16. A first-hand account; the author notes it is personal opinion.
The setup
A Home Assistant instance that had grown to about 1,600 entities across 21 areas with years of inconsistent naming (misspellings, rooms that had been repurposed, generic switch names). The goal: clean it up, build dashboards a non-expert could use, and expose a sensible subset to Apple Home.
From adviser to operator. Home Assistant's built-in MCP server, set read-only, let the model see everything and give good advice, but the human still clicked through every settings screen. A community MCP server, ha-mcp, exposes 84 tools across the entity registry, labels, helpers, automations, scenes and dashboards, which let the agent do the work.
Scoped credentials. It ran under a dedicated Home Assistant service account with its own revocable long-lived token.
Current conventions. The same project's best-practices skill was loaded so the model worked from current Home Assistant conventions, not stale training data.
The process: plan, then act
Plan as a reviewable document. The agent surveyed the live system and wrote a phased plan (clean data, then dashboards, then Apple Home). The author spent a day reviewing it with inline comments and questions before any change, much like reviewing a colleague's design doc.
The doc as running state. Each step had a checklist ticked as it completed, and every decision was written back with its date and reason. When a conversation grew too long, a fresh one was pointed at the doc and resumed from the next unticked item, with no re-explaining.
House rules written into the plan:
Full backup before each phase.
Tools auto-approved to avoid click fatigue, on the condition that the agent describes exactly what will change and waits for a yes. The author is explicit that a prompt-only rule is not a guarantee and that higher-risk work needs deterministic guardrails.
Pilot one device, check it together, then batch the rest.
Hard exclusions: never arm or disarm the alarm, open the garage or open a sprinkler valve without specific consent.
Results in four days
39 devices and about 340 of their entities renamed to a convention; 32 of 33 relays reclassified as lights and fans.
14 cameras and 399 entity IDs renamed; dependent automations updated, and older ones rewritten so future renames won't break them.
Everything sorted into areas and labels; a new room-by-room dashboard.
A latent bug found: a server-room temperature alert was sending to a phone that no longer existed, so it would have notified no one.
A slow air-conditioner status was traced to HTTP timeouts; moving one unit to MQTT cut status updates from 95 seconds to under a second.
Where it went wrong
Confident wrong diagnosis. It said the door locks had dropped off Wi-Fi; they deliberately run on Thread with Wi-Fi off. Acting on it would have wasted a night.
Hidden dependents. The renames broke a separate energy agent with 61 hard-coded entity IDs, 33 of which no longer existed. Once given repo access, the agent diagnosed it, opened a pull request with fixes, and added a new rule to the plan: check that repository before renaming. The mistake did not recur.
Taste stayed human. Dashboard design changes and rejecting a too-busy home page were the author's calls.
Lessons for engineers
Read-only access makes an adviser; scoped write access makes a co-worker. Give write access through a dedicated, revocable identity.
A plan document that holds checklists and dated decisions does double duty: it is the agent's resumable state across sessions and the human's audit trail.
Pilot-then-batch catches convention mistakes on one item instead of hundreds.
Rename-style changes need a dependency search outside the system being changed; external consumers of IDs break silently.
Treat the agent like a fast, well-read junior: approval gates exist so its confident errors are caught early and written down.
Review checklist
Does the agent act through its own revocable credentials, with hard exclusions for physical or irreversible actions enforced outside the prompt?
Is there a reviewed plan before the first write, and does every decision land in it with a date and reason?
Is there a backup or rollback per phase?
Do batches start with a single pilot item?
Before renaming or deleting identifiers, has every external consumer (other repos, agents, scripts) been searched?
Related entries
R112 (agents leaking data where boundaries were assumed but not enforced) and the project-memory entries under index subtopic 4.1.