Context engineering

The term gained currency in mid-2025 as agent builders argued that "prompt engineering" undersold the problem: in a multi-step agent, most of the tokens the model sees are not written by a human at all but assembled by the harness from tool definitions, retrieved documents, tool outputs, memory files and prior turns. Anthropic's engineering post "Effective context engineering for AI agents" (29 September 2025) frames the goal as finding the smallest set of high-signal tokens that maximises the likelihood of the desired outcome, treating context as a finite resource with diminishing returns.

The underlying constraint is that model performance degrades as context grows, even well inside the nominal window. Anthropic calls this context rot; related measurements include the "lost in the middle" effect and long-context benchmarks that show effective length below advertised length. Irrelevant or contradictory material also distracts the model, and stale tool output can be mistaken for current state. Cost and latency scale with input tokens on every step of an agent loop, so an agent that re-sends a growing transcript pays quadratically over a long run unless prefix caching is exploited.

The main techniques are well defined. System prompts should sit at the right altitude - specific enough to guide behaviour, general enough not to hard-code brittle rules - and be organised into clearly delimited sections. Tool sets should be minimal and non-overlapping, with tool outputs designed to be token-efficient (paginated, truncated, filterable). Just-in-time retrieval keeps lightweight references such as file paths, queries or URLs in context and loads content only when needed, instead of pre-loading everything. Compaction summarises a conversation nearing the limit and restarts with the summary; a lighter form clears old tool results that will not be needed again. Structured note-taking persists progress, decisions and to-do lists outside the context window, to be re-read later. Sub-agent architectures give focused tasks to helpers with clean contexts that return only condensed results to a coordinating agent.

Ordering matters for both quality and cost: stable material (system prompt, tool definitions, reference documents) first so that prompt caching can reuse it, volatile material last, and the current question near the end where recall is strongest. Context engineering also has a security dimension that prompt engineering lacks. Every automated source - search results, fetched pages, repository files, other agents' outputs, persisted notes - is a channel for indirect prompt injection, and memory that persists between sessions can carry an injected instruction forward. Provenance labelling, separating trusted instructions from untrusted data, least-privilege tools and review of what is written to long-term memory are part of the discipline, not an add-on.