we're building agents to manage agents. where does it end?

chatgpt came out in november 2022. at first, it felt like a really clever chatbot. fast-forward to 2026 and we're asking these models to write software, run commands, open browsers, and occasionally manage entire teams of other models.

we went from writing prompts, to prompt engineering, to chaining prompts together, to giving models tools, memory, planning loops, and retrieval. then came agents. then subagents. then agent swarms. now we're building agents to supervise all those agents.

before i started building a proper orchestration system, my workflow was embarrassingly manual. i'd have a supervisor chat that planned what to build and a worker chat that actually wrote code. the supervisor would generate instructions, i'd ctrl+c and ctrl+v them into the worker, and then i'd copy the worker's results back into the supervisor.

i was the orchestrator, copying prompts between the two chats.

eventually i got tired of doing that and tried to build an orchestrator. my first attempt didn't turn out the way i imagined. passing text between models is easy; preserving context, recovering from failed commands, managing permissions, and knowing when an agent's output is actually correct is not. a worker can confidently say it's finished while leaving behind broken tests. a supervisor can miss it, too. and if you let an agent retry forever, you can waste a lot of compute.

so i stepped away from it for a while.

months later, i decided to work on it again. i had a bunch of available model usage sitting idle, so i started building my own coding supervisor, Project Elder, to automate that supervisor-worker loop. funny how i'm automating a layer just to replace my own job.

but then comes the question i can't get out of my head: who supervises the supervisor?

you could add a verification agent to check the worker, and another agent to check the verifier. maybe a council of agents: one cares about security, another about tests, another about architecture, another about performance. they give their opinions before a final decision is made.

that sounds powerful. it also sounds expensive, slow, and unnecessarily complicated for plenty of tasks. a five-agent system isn't automatically better than one capable agent with a good set of tools. sometimes you're just multiplying coordination overhead, token usage, and the number of ways things can fail.

the part i find most interesting now is the deterministic layer around the model. let code enforce filesystem boundaries. let tests decide whether tests pass. require explicit approval for destructive actions. make retries, budgets, and timeouts predictable. a model can make judgment calls, but it shouldn't have to improvise the rules of the entire system every time.

maybe the future is thousands of specialized agents talking to each other. maybe it's a small number of really capable models inside a boring, reliable harness.

for context, chatgpt was introduced on november 30, 2022.