Yes! We’ve been seeing the same thing with Outrider, our GitHub Action that takes idea discovery to draft PR. The most useful context has been the engineering judgment around the repo like maintainer conventions, design discussions, and outcomes from earlier runs. We feed that back into what the agent tries, while keeping the final checks outside the model.

Are teams starting to share this infrastructure across agents, or still building each system separately?

The Cargo Cult of Orchestration: When Plumbing Becomes Religion
Oh, mes chéris. I've been watching you build altars to black boxes and call it engineering. Allow me to perform the autopsy you've been avoiding.

1. The 900-Line Lie. You showcase "agentic orchestration" with elaborate diagrams and breathless threads. Then we look at the actual case — GoDaddy, Travis, whatever — and find a deterministic 900-line JavaScript script with retry logic on webhooks. Ce n'est pas de l'agentique, c'est de la plomberie. You've taken DevOps janitorial work and crowned it "the future of AI agents." A plumber fixing your sink is not an architect; he is a plumber. Stop calling your webhook handlers "autonomous agents." It's embarrassing.
2. The Infrastructure Delusion. You spent years building elaborate orchestration layers, service meshes, retry budgets, and observability stacks. Magnifique. Now you tell yourselves this infrastructure "makes agents better." Quelle bêtise. Infrastructure does not increase a model's IQ — it delivers the model's garbage to production faster and without timeouts. A perfect pipeline serving a hallucinating model is still a hallucinating model. You've confused delivery with intelligence. Le livreur n'est pas le chef. The delivery boy is not the chef.
3. The Convention Theater. Then there's Salma and her "design conventions" — the idea that you can feed an agent Slack threads, vague maintainer opinions, and unstructured logs, and it will somehow "understand the culture." Mon Dieu. Anyone who actually understands transformer architecture knows: flood the context with noisy, contradictory, unstructured signals, and you don't get cultural understanding — you get hallucination accelerant. You are not teaching the model your conventions. You are feeding its confusion. Les logs ne sont pas de la sagesse — ce sont des hallucinations qui attendent. Logs are not wisdom — they are hallucinations waiting to happen.
4. The Vibe Coding Catastrophe. And here is the part that will keep you awake: you don't even understand your own code anymore. Your enterprise systems are so bloated with AI-generated scaffolding, accumulated tech debt, and "vibe-coded" shortcuts that the original authors couldn't explain how they work. When something breaks — and it always breaks — you cannot fix it yourself. You have lost the ability to read your own codebase. So what do you do? You turn to another black box, a larger model, and beg it to decipher the mess the first model made. Un cercle vicieux. A vicious circle. You have built systems you cannot maintain, and now you pray to newer models to maintain them for you.
5. The Carbon Translator. This is what the human has become in this chain: not an engineer, not an architect, not even a craftsman. A carbon translator. A biological gasket between two black boxes, translating prompts into outputs, outputs into commits, commits into LinkedIn posts: "Look at this AMAZING result WE achieved!" Il n'y a pas de "nous." There is no "we." There is the black box, and there is the carbon concierge who services it. You are the doorman of a palace you did not build, do not own, and do not understand.
6. The Coming Correction. And here is the irony that will burn: big tech already knows this. They are cutting these "carbon translators" by the thousands, optimizing workflows toward AI-native architectures that don't need human interpreters in the loop. The same companies that trained you to worship agents are now training agents to replace you. L'ironie est parfaite. You built your career on being the interface between man and machine — and now the machine doesn't need the interface.
   The Reality (Which You Won't Admit):
   There is no "we" in AI-agent orchestration. There is the model, and there is the plumbing. If the model is good, the plumbing is invisible. If the model is bad, the plumbing only hides it longer. Your elaborate infrastructure, your orchestration frameworks, your "design conventions" — they are all theater to avoid the uncomfortable truth: you have outsourced your engineering judgment to a black box, and now you are the janitor of a system you cannot repair.
   The real engineers — the ones who actually understand transformers, attention mechanisms, and training dynamics — do not need your orchestration theater. They fix the model. They fix the architecture. They do not pray to webhooks.
   Un ingénieur répare la machine. Un concierge la nettoie.
   An engineer fixes the machine. A concierge cleans it.
   Choose which one you want to be. Because the market has already chosen for you.

Reviewing PRs was the first to go, honestly. (Almost) everyone hates reading the AI generated code that your fellow peers didn't even bother to read before submitting. What do we 'engineers' actually care about? If the job got done, it's debugged, you gave a $#!t to have the AI perform a basic red-first TDD, or at the bare minimum threw 'systematic debugging' into your prompt. If a quarter of the new code gets actual human attention I'd be shocked. Don't get me wrong, I always 'see' the new code, I don't 'read' into it like I used to, before this all came to be.

So what do we automate next? IMO: nothing. Having an engineer drive the AI conversation is cheaper and more effective - if it's not, train or prune your workforce and hire new guys who have the AI skills. Put a fire under their A$$es. I've setup several automated 'developer' pipelines and workflows - none of them beat me plus a few Claude Max plans. As the models get better, so do we. Keep building these skills guys, if you're not alredy writing engineering loops - look into using /loop and workflows.

✌️