This piece is really good:
> The cost of building has collapsed, but the cost of aligning organisationally has not. If anything, it's gone up. When three different teams can each produce a working solution to the same problem in the time it used to take to write a proposal, the bottleneck moves from engineering to coordination.
We're still figuring out how to productively use coding agents as individuals, the next challenge is figuring out how to productively use them within teams. Coding agents reduce one bottleneck - producing working code - but that just moves the bottlenecks elsewhere.
(Note I said "working" code and not "good" code, that's a whole other thing.)
I still don't understand why coding agents are silo'd, and chat history is treated as disposable. Everyone on the team should be able to see all conversations and drop in and steer agents at any time, and chat history should be part of organizational memory.
I thought labs would have pushed collaborative steering by now, but I guess people got so TUI pilled they haven't even considered the possibility.
This is essentially what Shopify did with River (https://shopify.engineering/under-the-river), and we built out something similar at our company.
I'm sure the labs are working on this sort of thing - Managed Agents are probably the closest?
Yeah, I see the chat transcript as an important part of the work, and worth recording.
I've taken to linking to mine in commit messages, e.g. this one: https://github.com/simonw/simonwillisonblog/commit/e781e4eef...
(My favourite feature of the new Codex desktop app is that it has a comprehensive "Copy as Markdown" feature, which I can then paste into a Gist.)
I built this at my work and it didn’t turn out too well.
It’s a distributed agentic system on Temporal where all inputs are signals to the workflow. And then each agent has its own GNOME desktop either in a K8S pod or KubeVirt VM.
The biggest problem was context and ownership. The more we kept steering the model the worse it got until eventually it couldn’t complete its task. And on the ownership side, no one clearly owned the outcomes so it was just there… generating slop.
What will chat history give you that the output of that chat won't?
Was a specific piece of functionality intentional or an unintended byproduct of a previous change? Did someone else remove something but forget to take this out, or is it important? Etc
The same kind of value as a good commit history over just the finished code.
But you still have the commit/pr history.
Not sure what adding the ramblings with an LLM will add.
Context? Without it, it would be like starting a new session for each follow-up prompt
One mitigating factor is the increased productivity leads to consolidation, aka layoffs, meaning fewer people to align with. (Leading to further increased productivity, more consolidation, and so on ... Whether this is a virtuous cycle or a vicious cycle depends on perspective).
The vibe I got from the article was that now that technical work is faster, bosses expect everything else to be proportionally equally accelerated, even though LLMs either don’t help at all at those workloads or actively make things worse.
Different teams could already step on each others' toes before AI. If there is confusion over which team ought to develop something it might indicate an organizational problem. Crucially, if you ship something you must be willing to maintain it.
I don't think the author would dispute that. But the problem is the foot is bigger and the stepping is faster.