starting from here
I've been saying context is the entire engineering surface since before "context engineering" was a phrase. Then the phrase showed up and it meant something else: put the right documents in the prompt. That's not it.
Here's what I mean. The model is the thing you run on. Context is the program. If that's true, then context should be programmatic and just-in-time: assembled per turn, by code, from live state. Not authored ahead of time. Not retrieved and stuffed. Not a transcript.
The transcript is the problem
Most of what gets called an agent is a history of inputs and outputs with a loop wrapped around it. Every turn appends. When it gets long, something summarizes it. The "agent" is the loop, the state is the transcript, and all the ceremony (planners, reflection steps, memory modules, retrieval) exists to work around the fact that the context never evolved past a chat log.
Developers know what their customers need. They don't need an agent to search the entire space for it. But they also don't have to prescribe it. They can program their context to surface exactly what's necessary to get to the desired output in the shortest number of steps.
That's the whole question:
What program causes the model to produce the desired terminal output in the fewest steps?
Every extra round is the context program failing to surface something it could have known. "Agentic" is what you fall back to when you couldn't program it.
What that looks like in code
This is what proompt is. A block is a thing that fetches live state and renders it. A wall is a program that spawns blocks and assembles them per turn. A strategy is an ordering with cache breakpoints, because what's stable goes first and what changes goes last, and the model provider bills you for the difference. A cheap gate turn decides what persists into the expensive one. A block can run its own loop and hand control back.
None of that is a transcript. The conversation is one block among a dozen, and it's not even the biggest.
ByteBot is the wire under it. A Hub that executes and routes, and satellites that render events and send input. It's boring on purpose. The wall gets to be the application because everything around it is dumb.
Receipts
The framework runs the manager inside Indepreneur's product. A real company, real customers, ARR. Two services: the Hub with the wall, and a browser that is just a satellite. Multi-tenant on scope from the first line. Per-user tools narrowed from one stable toolset. A coaching pipeline that reads call transcripts and writes sprint plans with no chat surface anywhere. Every gap production found became a change to the framework, not a workaround in the product.
It also runs an intern in a Discord full of people who hate bots. He mostly doesn't answer. That's a context decision too.
Starting from here
The 2025 essay stays up as an archive. Everything it said about tokens, markets, and licenses is void. Everything it said about the framework happened. Three AIrrows Capital is a functional parody of a hedge fund and the lab where this stack gets tested in public; the founder hasn't clocked in yet.
This is a narrow opinion about where LLM ops goes. I think context is the whole job. The rest of the site is the evidence.