Context is a finite resource

For two years, everyone talked about prompt engineering: writing the instruction well. With agents that work through dozens of steps, the question has changed scale. At each step, the model receives a whole package: the system prompt, the definition of each tool, the entire history, every tool result, the retrieved documents. What decides quality is the composition of that package. The term now is context engineering.

More context, less attention

The Beginner level showed that the middle of a long text is read less well. Measurements published since go further: on many models, performance degrades as the context grows longer, including on simple tasks and well below the advertised limit. A large window is a capability, not an invitation to fill it. The right goal is the opposite: the smallest set of information that is enough to do the next step well.

What inflates an agent's context

• Raw tool results. A search that returns a 40 KB JSON of which three fields are used, an entire web page with its menus and footers.
• Tool catalogues. Connecting ten MCP servers often means fifty or a hundred tools whose descriptions take up thousands of tokens on every call, and a model that hesitates more between neighbouring tools.
• Failed attempts piling up, files read three times, endless command outputs.

What it costs

The "After launch" course put a figure on it: the context is sent again at every step. A forty-step agent whose context swells pays forty times over for the same tool results, with latency growing accordingly. Quality, cost and speed all point the same way here: trimming helps all three.

The first levers

• Tools that return little: the useful fields, pagination, a summary rather than the full object, error messages that say what to do.
• Few tools, clearly distinct: if a human hesitates between two tools when reading their descriptions, the model will hesitate too.
• References rather than content: keep a file path or a document identifier, and only read it again when it is needed.

Key points

• What the model sees at each step decides quality: that is the subject of context engineering.
• Performance drops as the context grows longer, well before the advertised limit.
• Concise tool results, a reduced tool catalogue, references rather than content.