Plenty of motor starter panels have a Hand-Off-Auto switch on the door.
Hand means a person runs the motor. Auto means the controller does. Off means nobody does. The switch picks who’s driving. It doesn’t pick which safety rules apply. On a panel wired right, the overload and the e-stop sit in the circuit no matter where the switch points. Hand doesn’t get a shortcut around the interlock because a person is standing there.
In my sketch of that panel, someone has scratched a tiny star next to the Auto position.
Agents are a third position on that switch. Dropbox just showed what wiring it right looks like.
Two drivers, then three
Dropbox engineers Greg Unrein, Josh Jensen and Christopher Wildman wrote up how they made Reclaim AI-native without starting over. Reclaim is the Dropbox-owned calendar assistant that finds and moves time for tasks, habits and meetings.
Before the rebuild, a calendar could change two ways. The user made a change by hand, like moving an event. Or the automated scheduler adjusted blocks in the background as the day shifted. Hand and auto.
The agent is the third driver. A user describes what they want, and the agent works out what to do. The post is honest about why that’s harder: a request like “make time for this tomorrow” could mean finding an open block, creating an event or moving something that’s already there. Some of those touch other people’s calendars.
One interlock for everyone
Here’s the decision I like. Reclaim represents each scheduling capability as a Schedule Action, its internal name for a standard operation like creating or updating an event, changing an RSVP or finding availability. Per the post, whether a request comes from a user, the scheduler or an agent, it uses the same Schedule Action type, with the same validation and commit process.
They say plainly why they didn’t build an agent path on the side:
Creating a separate AI-only path would have duplicated functionality and made Reclaim’s behavior harder to keep consistent across interfaces.
That’s the interlock wired in series with every position on the switch. Moving one meeting can change someone’s availability and shift flexible events elsewhere. If three implementations each enforced those rules, they would drift. With one, engineers change an operation in one place.
Preview before it goes live
The second piece is Preview Mode. It gives the user a temporary version of their calendar where they can review what the AI suggests, see how a change ripples through the rest of the schedule, and check the effect on shared events before any updates go out to other attendees.
To make that fast, they reworked the scheduler as a pure function. It can calculate a proposed schedule without saving anything to the real calendar, and it reruns every time the user adjusts the preview. A copy of the calendar data sits in Redis so the recalculation has what it needs.
On a test floor that’s a dry run. You step the sequence with the outputs disabled, watch what would have happened, then enable outputs. Same idea as keeping the sandbox as the fixture, not the record: the copy is where you try things, and the real calendar only changes on commit.
A few other choices from the post are worth noting. Each request gets only the calendar context and tools it needs, not everything at once. They built their own agent loop and connect directly to model providers rather than adopting a broad framework, and they’re upfront that this means more code to maintain. And Reclaim speaks MCP both ways: an MCP client inside its agent loop, and an MCP server that exposes selected Reclaim tools to outside AI clients, which call the same tools.
Where I’d put the interlock in your stack
Most teams adding an agent do the opposite of Reclaim. The agent gets its own API key, its own route to the backend and its own looser rules, because it was faster to ship. That’s an agent position on the switch with no overload in the circuit. It’s also how you end up asking who gave the agent a badge.
The Reclaim pattern generalizes:
- One operation per capability. Not “the agent version” and “the user version.”
- Same validation and commit for every caller. The caller identity changes what’s allowed, not which code runs.
- Preview on a copy. Anything that touches other people goes through a dry run a human can read.
- Commit is a separate step. The agent proposes. The commit path decides.
At the network layer, the same idea is why I like putting agent tool calls through one gateway that applies the same authorization policy to every caller. That’s what agentgateway does for MCP and agent traffic, and it’s the same instinct Reclaim applied inside the app: one path, one set of rules.
I work at Solo.io, which makes agentgateway.
And design the preview for the person reading it. As I wrote in human attention is the spec, attention is the scarce part, so the preview should carry what they need to decide quickly.
Pick the driver with the switch. Wire the interlock so it doesn’t care.
Sources
- Evolving our calendar assistant Reclaim to be AI-native without starting over, Greg Unrein, Josh Jensen and Christopher Wildman, Dropbox, September 29, 2026.