The model is the commodity
#8am-ai#context-engineering#deep-dive#evals#project-state
David Olsson
The agentic thread is the corpus's clearest arc. The context thread is its clearest conviction.
The conviction: the model is the commodity, and the context is the lever. What separates a useful interaction from a useless one is rarely the model. It's whether you can hand it a structured, current context โ and whether you understand the work well enough to know when the output is wrong.
The vocabulary changed across two years. The conviction didn't.
2024: scrubs
The earliest name for it is scrubs โ instructions and metadata you attach to an LLM so it does a task reliably. The move in April 2024 is already this: don't reach for a bigger model, give the model a better frame.
By August the group runs AI over Hanif's spreadsheet data. Not to replace the analysis. To extend what's already there. Context as raw material, not magic.
A model with a good instruction set beats a better model with none.
2025: corpus, portability, the limits
The conversation gets more demanding. It's no longer "write a scrub." It's "build and maintain a corpus."
The corpus is the asset. The freshest, best-structured context compounds. The model subscription doesn't. People start treating their documents, notes, and prior outputs as a substrate the model reads.
Portability is the bottleneck. Context is trapped at the app level. You build something good inside one tool and lose it the moment you switch. The group rates the value of portability as extremely high because it's so hard to get.
AI gets you 95%. Scott's line about his investor table: the model gets him most of the way, the last 5% is human. The gap isn't capability. It's the context and judgment only the human holds.
2026: how do you know
By 2026 the conviction has a sharper edge and a name for its missing piece. The group keeps circling one question: how do you know? Not "is the output good." What in your process tells you it's good, independent of the thing that produced it?
This is context engineering grown up. The point is no longer feeding the model better inputs. It's building an external check on a self-contained process. A workflow that grades its own work has no observability into its own reinforcement.
The project-state pattern is the structural answer: a corpus that's also a map, so both human and model can trace a claim back to where it came from.
The foundational-knowledge worry lands in the same place. Juan's 2026 framing: you have to be able to combat an AI suggestion โ to critique an approach โ which you can only do if you understand the domain. The context isn't only what you give the model. It's what you carry yourself.
Three claims never moved across two years. Reach for context before capability; almost every "the model can't do X" was "the model wasn't given enough to do X." The corpus compounds and the subscription doesn't. Someone has to be the eval.
What's still open is portability โ context that travels, inspectable and current. The group has the conviction and the patterns. The clean answer to portability is the next piece of the substrate.