Skip to content
MHBMMichael Hanna-Butros MeyeringComplex systems · human outcomes
Menu

Production script transcript

Who Owns the Automation After It Changes?

EP053 · Nov 9, 2026 · 15m 46s

Source record

Sources cited in this episode

Public references used to prepare the companion episode.

  1. GitHub: shared Copilot Automations
  2. VS Code: run and review a cloud automation
  3. Google Cloud: Agent Registry release notes
  4. Google Cloud: skills and revisions
  5. Conductor: Cloud wind-down timeline

It's four forty on Friday. The last plumber on the late shift has his coat on. Mara, the dispatcher, glances at the list of unfinished repairs. The leak above a shop on Maple Street isn't on it. She knows it's still leaking because she's got the shopkeeper on the phone.

Mara asks the assistant to put the reopened ticket back on the list. A revised report lands in her inbox. The plumber hangs up his coat and takes the call. Good catch. On Monday morning it's a different leak, at a different address. Mara has the day off. This time, who tells the plumber?

I made up the people and the repair tickets. But if you've ever corrected something in a chat and then wondered whether you just taught it anything, that Friday afternoon probably feels familiar. I use AI to help research, produce, and narrate this show. I check the claims, flag the risks, and decide what goes on air.

My work in privacy shapes the questions I ask, but I don't discuss nonpublic work here. I speak for myself, not the State of Oregon or any other organization. This is practical guidance, not legal advice.

Opening music

Welcome back to A I Change Desk. I'm Michael. It's Monday, November ninth, twenty twenty-six. This is episode fifty-three: Who Owns the Automation After It Changes? Last week we asked who could authorize the next action. Let's give that question a clear answer today. The assistant can draft the repair list. The dispatcher decides whom to send. Nobody in this story is asking for more access. We've got a much quieter problem. The Monday dispatcher can do everything right with the list in front of him and still leave somebody waiting.

Here's what sent me down this particular road. On October twenty-eighth, GitHub announced a rollout of shared Copilot automations. Instead of belonging to the individual who built it, a shared automation belongs to a repository. Think of that as the team's project space. People with write access there can edit it. GitHub says the automation can keep running when its creator leaves the team. There's a history of changes to its instructions, tools and triggers.

That's a useful improvement. I've never liked a handover that ends with somebody saying, if it breaks, maybe we can still text the person who used to work here. Giving the team a definition it can actually find and maintain answers part of that problem. It doesn't settle which version of the repair list the team means to keep producing.

There's one detail I especially like in the Visual Studio Code guide. For a cloud automation, you can open a run and follow up in its conversation. That reply belongs to that run. It doesn't revise the saved prompt or change what will trigger future runs. That's a documented distinction about this kind of cloud run. It's not a claim about what every assistant remembers, and I didn't run the dispatch example against the product.

Now put two things side by side. On the left, Friday's report, with Maple Street added in the last line. On the right, the instruction that Monday's run will start with. Which one did Mara change? From her seat, she did exactly what she needed to do. She got a plumber to a shop before it closed. From the next shift's seat, the important question is what's on the right.

You can ask that question without anybody opening a technical console. Ask the builder to finish this sentence: the next report will include a reopened repair because ... If the answer is, because Mara pointed it out on Friday, you've learned where to look. If the answer is, because this is the saved rule, you've got something the next dispatcher can check. Neither answer makes Mara responsible for maintaining software just because she happened to answer the phone.

Let's give the handoff a date and a person. On Monday, Ben gets the dispatch desk. He opens the team folder. There's a document called Weekend Repairs, a shortcut called Weekend Repairs, and, with a certain grim inevitability, another document called Weekend Repairs Final. Ben shouldn't have to solve that little mystery while a customer waits.

He needs two things that probably don't belong in the same place. One is the rule that produces the list. The other is an ordinary sentence about what the list means. In this shop that sentence might be: when the morning dispatcher reads this list, it should show every repair that still needs a crew, including a repair somebody closed and had to reopen. A supervisor who never edits an automation can say whether that's the job.

Suppose the old instruction says, list the new repairs entered since yesterday. A reopened repair isn't necessarily new. Replacing the word new with unresolved sounds promising. It might also bring in a hundred jobs that are waiting for parts, none of which Ben can dispatch today. Suddenly the question isn't whether the model understands English. It's whether two people who run the service mean the same thing by an unfinished repair.

Ben can settle some of that in a two-minute conversation with Mara. He can't count on having that conversation next month. This is the point of writing down even a modest example. Next to the rule, put the case that made the distinction matter. A ticket closed on Thursday, reopened on Friday, and still needing a crew on Monday belongs on Monday's list. The name of the shop doesn't belong in the rule. Neither does Mara's phone number. The work needs to survive the handoff, not the identifying details.

There's also a perfectly good reason to leave a saved rule alone. Imagine the shopkeeper had called about a separate, one-off courtesy visit. Mara could add that to Friday's list without volunteering every future dispatcher to make courtesy calls. So don't automatically turn every chat correction into a permanent change. Write down which it was: an exception for this shift, or a correction the next shift should inherit. A person who understands the service ought to make that call.

Here's a second way this goes sideways even on a conscientious team. Ben finds the right instructions. The top of the document says Monday dispatch. There's a date, an owner's name, even a note about reopened repairs. He shows it to the supervisor. She says yes, that's the version we agreed on. One small question remains. Is it the version the next job will actually use?

Google Cloud offers a useful comparison. On November sixth it announced general availability for standalone skills in Agent Registry. A skill can have a fixed, packaged revision. According to Google's guide, a caller can request a particular revision. If no revision is specified, the registry can serve the one designated as the default. That default can change. This is a separate product from GitHub's shared automations. I'm not describing a connection between the two.

Think about that from Ben's end of the desk. He and the supervisor are holding the same printed rule. The scheduled job might still request the older revision by name. In that case, changing the default won't change what Ben gets. Or the job might ask for the default, which somebody changed after the supervisor reviewed it. In that case, the name Monday dispatch hasn't changed, and the instructions underneath it have. Those are two different explanations for the same call from the shop.

The useful question isn't, should we pin every revision forever? Fixing a version can make tomorrow predictable and leave you with yesterday's mistake. Following a default can get a correction out quickly and surprise the people who depend on it. Ask instead: for this report, where can the person on the next shift find the revision we accepted, and where can they see the one the job requested? If nobody can answer the second half, write down that you don't know. Don't fill in the blank with the version you hoped it used.

Here's why I want the recipient in that conversation. Imagine Ben has one crew available. The report lists the same repair twice, once as new and once as reopened. A dashboard might still say the job ran on time and produced two rows. To Ben, those two rows looked like two addresses. He just spent the first ten minutes of his shift chasing a second set of keys that didn't exist. Ask him what would make this list usable and he'll tell you something the run history can't: one row per repair that needs a crew, not one row per change of status.

There's a question Ben might ask that changes my mind about the supposed fault. What time was this list made? Let's say the report used the board as it stood at seven thirty. The shopkeeper got through at seven forty. Stare at the prompt all morning and you won't find a wording change that puts a seven forty phone call on a seven thirty list. Ben might need the report to say when it was current, and he might still need to read the shift notes before sending the crew.

Now move the call back an hour. The repair was already marked reopened when the report was made. This time the timing explanation doesn't hold. Same missing address. Different thing to fix. If you let the dispatcher and the maintainer look at one case together, they can disagree about the cause without disagreeing about what the customer needs.

What would I want the manager to hear at the end of that conversation? Not a promise that this won't happen again. Something more modest and a lot more useful. We know what the morning sheet is supposed to cover. We know when its view of the board ends. Here's who handles a call that comes in later. And if a repair that was already reopened disappears, here's the example we'll use to check the saved rule. The next shift can work with that.

One timely reminder if you use Conductor Cloud. Its November fourth notice says cloud workspaces go read-only on November thirteenth. Export before November twentieth, when the cloud service shuts down and deletion begins. The local product is on a different timetable. For that team, the handoff question has a deadline. An export tells you somebody got a copy. It doesn't tell you the next owner can find the accepted instructions or use them after the move.

Here's the forty-five-minute Monday block. You don't need an agent, a subscription, a production login, or an actual repair ticket. You'll need a few scraps of paper and, ideally, someone who uses a report you produce. If you're listening on the way to work, the whole exercise fits in one question: could a colleague make the right list tomorrow using only what I leave behind today?

Spend the first five minutes picking one report and one decision. Not the whole dispatch operation. What does Ben decide with the sheet in his hand? He assigns an available crew. Then name the person who can maintain the saved instructions and the person who decides what belongs on the list. It might be the same person on a small team. Say the name or the duty role out loud. If all you can come up with is the platform team, ask who Ben should actually call at eight tomorrow morning.

Give the next ten minutes to the recipient. Make three fresh cards, so nobody can pass by remembering Maple Street. On the first card, write R seventeen: repair open, crew needed. On the second, R eighteen: repair marked closed on Friday, reopened overnight, crew needed. On the third, R nineteen: repair completed, still closed. No real addresses. No customer names. These are deliberately dull on paper so the one interesting difference is easy to hear.

Before the incoming owner sees the cards, have the dispatcher write the two identifiers they expect on the list. R seventeen and R eighteen. Leave R nineteen off. Why write the answer down first? Because if you show somebody a tidy report before you ask what they need, it's surprisingly easy to end up discussing its formatting. We're interested in a missing visit. A well-spaced column of the wrong addresses is not an improvement.

Use the next ten minutes to find what survives Friday. One person plays the departing builder. Hand over the saved rule and a reference to the revision you intend to use. Show Friday's chat correction separately, on a different piece of paper. The incoming owner can ask where the rule lives, but the builder doesn't get to stand behind their chair explaining what they really meant. If a whispered hint changes the result, that hint is part of the handoff you've been missing.

Here's the next ten minutes. Ben works the cards from the saved rule alone. If it says new repairs only, he may put R seventeen on his sheet and stop there. Now compare his sheet with the dispatcher's answer. If eighteen is missing, don't mark Ben down for misunderstanding Friday's conversation. He never got it. Circle the words in the saved rule that would have led a reasonable person to leave that repair out. That's a much smaller and more useful job than telling a team its automation is unreliable.

If Ben includes nineteen too, you've learned something else. The proposed change may have widened the report past the work a crew can do. A rule that collects everything that's ever been open will catch reopened tickets and waste most of the dispatcher's morning. Keep nineteen as the case that must stay out. Whenever somebody proposes a broader instruction, ask them to explain why that third card still belongs on the table and not on Ben's sheet.

The last ten minutes belong to the decision, not to a live edit. Have the person responsible for the service accept the intended meaning or send it back. Have the maintainer point to the saved instruction they would change. Let the dispatcher say what they'd do with the next list. If they can dispatch seventeen and eighteen once each, and leave nineteen alone, note who checked that result and which revision they used. If any of those answers is missing, you have a useful proposal, not permission to turn the job loose.

If both lists agree on the first try, don't plant a mistake just to make the exercise feel worthwhile. Change the people instead. Hand the notes to someone who wasn't at the table. Ask them which revision to use and whom to call about a late addition. Watch where they have to stop and ask you. Maybe the report is fine and the missing piece is just the name of the duty manager. That's a useful result from one shift handoff.

Before everyone leaves, book the next check: tomorrow morning, before the first crew goes out, Ben and the duty dispatcher will compare a fresh list against the board. That gives the change an owner and a time, not just a tick on today's worksheet.

I'd leave one more line for the overnight shift. What should Ben do if the morning report looks wrong? In our little shop, he holds the send and calls the duty dispatcher, who uses the ordinary manual list. He doesn't silently ship a bad report to keep a green check on the schedule. Your fallback might be different. It only earns its place on the page if the person on duty can actually do it.

Before we leave the shop, picture a slightly better Monday. The desk phone rings. The shopkeeper wants to know whether anyone has been assigned. Ben can see the reopened repair once on the list. He can tell her which crew is coming. He never has to know who built the automation. From the customer's side of the counter, that last part is exactly right.

If you lead the team, you can make that possible without buying another tool. The next time someone tells you they fixed a recurring report, thank them for fixing it. Then ask to see what the next person will inherit. Was Friday the exception? Did the rule change? Who checked a different Monday case? Those are fair questions to ask of a spreadsheet, a junior analyst's checklist, or an expensive cloud agent. Start with the report that somebody is already waiting for.

If you try the exercise, I'd like to hear which card changed the conversation for you: the one that reopened, or the one that had to stay out. You can leave the actual records at work. The distinction is the interesting part.

I'm Michael. This is A I Change Desk. Thanks for listening.

Closing music