- Author
- Chris Loy
- Published
- 3rd August 2025
- Kind
- Personal blog explainer with a worked toy example
- Original
- chrisloy.dev/post/2025/08/03/context-engineering
- Subject
- Harnesses and agent loops
- Themes
- What lasts, Getting good
What it is
Chris Loy writes a series called Machine Learning for Engineers. This post explains the shift from prompt wording to managing every token in context. It ends with four rules.
What it says
- Loy says prompt engineering was often trial-and-error guesswork, closer to incantations than engineering.
- The post says context now holds examples, tool calls, documents and memory. That strains the window.
- Loy argues for treating the LLM as a skilled analyst to brief, not an oracle.
- Loy calls RAG one design pattern for context. Others include tool calling and context compression.
- The post shows a UK box office question. It needs the date, a document and a calculator tool.
- Loy says multiple agents hand off token sequences. The hand-off should be treated as an API contract.
What evidence it gives
The post gives a toy walkthrough. A model quotes an outdated UK box office average of 24 million pounds. With context added, it gives 18.8 million pounds per week. No benchmark or measurement is reported.
Interests and limits
The post is a personal explainer, not a study. The film and box office examples are illustrative. The design pattern analogy is the author's framing. Nothing was tested across models or tasks, and no outcomes were measured.
Connections
These links are this wiki's own reading, not claims the sources make.
- Layers of Memory, Layers of Compression. Kellogg reads multi-agent research as an exercise in compression. Loy says agents hand off token sequences, and calls that hand-off an API contract. Then a contract is a statement of what a compression must preserve.