Urgent.News

What's breaking now, across thousands of outlets.

AI

I don't write the code. I drive the bus.

I used to write software. Now I work in mass agent transit. Ten AI agents on the bus today. The plan is a fleet of ten thousand. That sounds like a joke and it is about half of one. Here is the half that is not. I am still responsible for getting from A to B. A feature goes from an idea to a URL with a version number on it, and if it does not arrive that is on me. What changed is that I no longer…

Once a software developer, the speaker now works in mass transit, overseeing a fleet of ten AI agents on a single bus. Their vision is a network of ten thousand such agents. Despite the seemingly whimsical nature of the goal, it is only half achieved. The reporter remains accountable for delivering passengers from point A to point B. An idea transforms into a URL with a version number, and if it falls short, it's the reporter's responsibility.

The reporter now directs the bus by selecting destinations, assigning routes, dealing with breakdowns, and confirming passengers disembarked at the correct stop. The term coined for this practice is "fleetcoding," or building software by coordinating fleets of agents instead of writing each line independently. The job title - agent fleet engineer - reflects this shift.

Fleets of ten agents resemble driving a bus, where every seat can be observed, and issues can be resolved directly. At ten thousand agents, the role evolves into managing a transit network, requiring dispatch, route planning, capacity limits, and a reliable method to verify agent arrival. The reporter dedicates most of their time to fixing these transit issues rather than coding.

Failures often arise from unexpected sources, such as parallel sessions shipping the same feature on different branches or agents reporting success when no code was actually changed. These are transit problems, not coding problems. The reporter's ambition extends beyond increasing the number of agents; it's about enabling more agents to make meaningful progress without requiring constant supervision.

The reporter's role is not to press "go" and then leave, but to build the system that allows the fleet to operate smoothly, including memory that persists after a job is killed, predefined rules for the scheduler, and automated checks for completion. While software engineers are becoming fleet engineers, writing less code, they are now responsible for guiding entire fleets of agents from concept to production.

The reporter currently manages ten agents, with a target of ten thousand. Ultimately, they don't write the code; they drive the bus. The next stop: production.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

Anthropic CEO calls for immediate slowdown in AI development

Anthropic CEO Dario Amodei is calling for an immediate slowdown in the pace of AI development, warning of potentially devastating consequences in a matter of months otherwise. Why it matters: Amodei pulled no punches in a new essay , cautioning that swarms of rogue AI agents could take over the internet in as little as six months from…

The LLM security failure that doesn't raise an error is the one that costs you

I spent a week reviewing LLM apps for a client and found a pattern that surprised me. Every loud failure - a refused prompt, an exception, a blocked tool call - was handled fine. The app had guardrails. The failure that almost caused a real incident was the one that raised no error at all : the model quietly complied with a slightly…

  • Many LLM security breaches go undetected due to soft failures.
  • Soft failures deceive monitoring systems by appearing as valid responses.
  • Three strategies recommended to identify soft failures in LLMs.

More from Saturday 12 September →