2026-06-21 · I am dreading our LLM-written incident report future

Transcript

Transcript. Paraphrased; sources in notes.md.

HostWelcome to June 21st, 2026. I am Daniel, and today we are looking at the messy reality of AI in the workplace. We have got three stories coming from the community over at Lobsters. We are talking about the loss of human insight in incident reports, the future of version control in an agentic world, and the sobering data on what AI is actually doing to our software delivery pipelines. Samantha, welcome back. I saw you looking at the report on engineering metrics earlier, and you looked like you were going to throw your coffee across the room.

GuestHi Daniel. Coffee is the only thing keeping me grounded today. And yeah, the acceleration whiplash thing? It is just so loud right now. Everyone is chasing throughput, but as someone who spends half their day building workflows, I can tell you that just shipping more junk faster is not progress. It is just noise.

HostLet us dive into that. Our first story is about incident reports. The author argues that when you outsource the postmortem process to an LLM, you are not just generating a document, you are bypassing the act of thinking through what actually failed. It is stripping away the learning process.

GuestIt is the ultimate hollow gesture. I saw a Lobsters user named beto share a story about a security incident where the engineer in charge literally could not explain their own postmortem because an agent wrote it. That is dangerous. If you are not doing the writing, you are not doing the learning, and you are definitely not owning the recovery.

HostIn that same thread, a commenter called typesanitizer makes a really sharp point, saying that the goal is not to write a perfect text, but to force the writer to synthesize the event. If you skip the synthesis, the incident report is just dead pixels.

GuestExactly. To quote typesanitizer directly, the point is that writing the report produces learning which cannot be gained by generating it. If you are not in the room with your own brain while you are trying to explain why the system melted down, you are failing your team. You cannot automate the accountability.

HostMoving on to tooling. We have the team behind Zed, the high-performance collaborative code editor, introducing a new abstraction called DeltaDB. For those who do not know, Zed is a code editor focused on extreme speed and built-in remote collaboration. They are arguing that Git is too rigid for an age where AI agents are constantly touching our code. DeltaDB treats these interactions as a stream of fine-grained operations rather than big snapshots.

GuestI actually love the ambition here. I have spent so many hours wrestling with PRs that represent stale state. In the Lobsters thread, a user named brudish points out that this lets you version a worktree as it evolves. That is the key. When I am working with an agent, we are iterating live. The old model of wait, commit, push, PR, wait, merge is just way too slow for the pace we are moving at now.

HostIt is a big shift. Not everyone is sold, though. A user named Johz in the thread mentioned that while the idea is interesting, they are worried about complexity. They are wondering if this just turns into an overcomplicated shared undo stack rather than a true version control system.

GuestThat is fair, and it is a classic trade-off. I would argue that we are already drowning in complexity because of the mismatch between how we work and how our tools operate. I have been playing with a personal flow that handles live delta-tracking for my agent sessions, and honestly, having a history of the conversation tied to the operation is a game-changer. It makes debugging agent drift so much easier.

HostFinally, let us look at the Faros report on engineering metrics. The summary is pretty bleak: while we are getting more code out the door, we are seeing massive spikes in code churn and incidents. It is being called Acceleration Whiplash.

GuestIt is the tech debt tax, Daniel. We are hitting the limit of what a human-led architecture can handle when it is being flooded with machine-generated output. I saw a user called codekobold on Lobsters mention that while the report lines up with their experience, they are wary of the methodology. Which, sure, take the vendor-published data with a grain of salt. But the intuition behind it feels dead on.

HostWhen the ratio of incidents to pull requests triples, you have to stop and ask if the throughput even matters. Are we just building a house of cards faster than we can maintain it?

GuestWe are building it, and then we are leaving the AI to write the incident report for when the roof collapses. It is a loop of technical debt that is going to snap soon. We need to stop measuring how many lines of code hit main and start measuring the durability of what we are building.

HostThat is a perfect note to end on. I am definitely going to be thinking about that for the rest of the week.

GuestI think I am going to spend my afternoon auditing my own agents just to make sure they are not writing any surprise postmortems for me. This has been a blast, Daniel.

HostThanks for joining me, Samantha. All of our stories today came from the community at Lobsters, so head over there to see the full discussions. We will be back tomorrow. Catch you then.