I'll bet a bottle of good whiskey that within a year someone will put "per-task agent context" on a pitch deck and the whole room will nod along without anyone asking what it means.
Every buzzword needs two people: someone to name it, and someone with enough reach to repeat it. The thesis of this post is that the second layer is already in place. Every day I spin up three to five sub-agents from my Claude Code with clean context. I still don't have a name for that practice that would pass through a PM. And that's why I'm betting I'll get one soon.
The bet
By spring 2027, the agent-dev crowd will settle on a folk term for task-level context isolation. I don't know the exact shape: "per-task context," "agent isolation," "context fork." But the shape will come. It will be short enough to fit next to a logo on a slide deck, and vague enough that everyone imagines something slightly different under it.
This isn't a technical prediction. It's a bet on how a meme is born.
Why this, why now
In one week of June 2025, Tobi Lütke drops a tweet saying "context engineering" is a better description of what developers actually do with LLMs than "prompt engineering." Six days later Andrej Karpathy amplifies the same term in a single tweet, defining it as the "delicate art of filling the context window with just the right information for each step." He didn't coin it. He made it resonate. Three months later Anthropic builds an engineering post on it: "Effective context engineering for AI agents." The term moved from Twitter to blogs and got a vendor stamp.
That's how the pattern works. A coiner from the second layer, an amplifier with reach, and finally legitimization by a vendor. So what's next? It's not better context curation. It's slicing the context up. When you have a long-context problem, you can polish it, or you can isolate it. Isolating is cheaper, and that's why it wins.
What's already running without a folk name
The cleanest proof that the practice is ahead of the language is the Claude Code documentation. On sub-agents it says: "All the intermediate noise stays inside the subagent's context and never touches the main conversation." One sub-agent can chew through tens of thousands of tokens and return a distillate of 1,000–2,000 tokens to the main conversation. It's a production feature, not a research preview.
Next to it, the Claude Agent SDK has an isolation flag, and the Claude API has beta strategies like clear_tool_uses_20250919 for context editing. LangChain added it in July 2025 as one of four patterns alongside write, select, and compress. The technology and the engineering are in place. What's missing is the slogan a person who has never opened a terminal can understand.
That's why I'm betting. The vacuum between "it runs in production" and "we know how to name it" is where buzzwords are born.
Where the ceiling is
I have the ceiling in my head. Chroma published work in 2025 on "context rot" across eighteen models, showing accuracy drops with context length, which is an argument that pushes isolation forward. But isolation itself has a limit. The Microsoft Azure SRE Agent worked nicely until it grew to fifty sub-agents, then it fell apart. Handoffs hold up for roughly four hops. Beyond that it's a telephone game, with sub-agents inheriting the mistakes of their predecessors.
I take that ceiling seriously. But a meme doesn't die on a technical ceiling. A meme dies on a better meme. And no one has offered a better meme yet. That absence is why I believe we'll get the label by spring 2027.
The prediction
The coiner won't be Karpathy. It will be someone from the second layer: a CEO or founder with their own product and a big enough Twitter, Lütke-type. Karpathy amplifies the term in a single tweet. Anthropic or LangChain ships a second engineering post that legitimizes it.
The term won't win because it's the most accurate. One of the versions will win. And whoever is already writing that practice into their own repository under a concrete name will, the moment the meme takes off, be a step ahead. Not as the coiner, but as someone who has something to show under someone else's label. I have my pick written in AGENTS.md at the root of my monorepo. I recommend you write yours down too, while it's still quiet.
