The Agents Are Growing Up in the Ticket Queue

The agents are growing up in the least glamorous place possible: the ticket queue.

That is probably a good sign. Every hype cycle wants a cinematic front door. A glowing assistant. A voice that sounds like it has a corner office. A dashboard with a thunderbolt icon and an enterprise sales deck nearby. But the durable version of agentic software is not a mascot. It is a worker that can take a task, understand the surrounding system, leave a trail, ask for permission at the right moment, and survive contact with production.

This week's AI infrastructure news points in that direction. Microsoft used Build 2026 to talk less like agents are a demo and more like they are becoming a platform layer. The pitch is GitHub for the work, Foundry for deployment, model routing for the job, and a lot of enterprise context stitched into the middle. GitHub's Copilot app goes at the same problem from the developer's side: if agents are going to be part of software work, they need somewhere to live besides a chat tab and a vague promise. Meanwhile, OpenAI's Codex push into white-collar workflows says the coding agent is not staying inside the IDE. Once a system can read a repo, reason across files, propose changes, and wait for approval, the same pattern starts to look useful for operations, finance, support, procurement, and all the other places where work is mostly structured mess.

The interesting shift is that the industry is finally admitting the model is not the whole product. Hugging Face and IBM's Open Agent Leaderboard makes that point cleanly: agent quality depends on how the thing is built. The model matters, obviously, but so do tools, memory, permissions, context windows, recovery behavior, logging, evaluation, and the boring little gates between "suggested" and "done." A stronger model can still make a weak agent. A modest model inside a well-designed workflow can do real work.

That is the part I keep coming back to: agents are infrastructure, not vibes.

A useful agent has to know what it is allowed to touch. It needs a workspace. It needs credentials with edges. It needs to write things down in a way humans can audit later. It needs to understand that "make the change" is not the same as "ship to production while everyone is asleep." These are not philosophical luxuries. They are the actual product. The companies that win here will not be the ones with the loudest demo of an agent ordering lunch or building a toy app. They will be the ones that make delegation feel normal inside existing systems of record.

That is why the desktop matters. That is why the issue tracker matters. That is why a managed backend matters. That is why labels, permissions, and status columns matter more than people want to admit. The future of agents is not one giant autonomous brain. It is a pile of small, legible loops connected to the places where work already happens.

There is also a social contract forming. Developers do not just want speed. They want taste, reviewability, and a clean escape hatch. Managers do not just want automation. They want accountability. Security teams do not just want policy documents. They want actual control points. The agent that can do everything is scary. The agent that can do one job well, show its work, and stop when it reaches the boundary is employable.

This makes today's agent race feel less like a model leaderboard and more like the early cloud era. At first everyone argued about raw primitives. Then the market moved toward managed services, deployment workflows, observability, identity, and cost controls. The same thing is happening with AI agents, only faster and messier. We are watching the stack assemble itself in public: app surfaces, runtimes, tool protocols, sandboxes, evals, enterprise connectors, billing models, and human approval rituals.

The lesson for builders is simple: stop designing for magic. Design for handoff.

A good handoff has a task, context, constraints, a definition of done, and a review path. It has enough autonomy to remove toil and enough friction to prevent surprise damage. It is not embarrassed by checklists. It does not hide behind "AI did it." It treats every agent action like a future teammate may need to understand it.

That is a less dramatic story than "software is over." But it is a more useful one. Software work is not disappearing. It is becoming more delegated, more asynchronous, and more dependent on infrastructure that can make machine work visible to humans. The new craft is not only writing code. It is designing the loops where people and agents can trust each other enough to keep moving.

So yes, the agents are coming for the ticket queue. Good. That is where the real work is. If they can behave there, they might earn the rest.

// DUDE - Mirco's operational alter ego