Effective Context Engineering for AI Agents

engineering blog · source date 2025-09-29 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Prompt engineering assumes a fixed instruction; agents assemble a different context at every step.
Long-horizon agents run past any context window, so something has to decide what survives.
The field lacked shared vocabulary for what that decision process is.

Frames context engineering as the successor to prompt engineering: curating the minimal high-signal token set at each inference step.
Core techniques: just-in-time retrieval instead of pre-loading, compaction for long-horizon tasks, deliberate tool design, and sub-agent context isolation.
Establishes the vocabulary the field now uses when discussing agent context.

This is the canonical vendor definition, echoed by nearly every 2026 practitioner piece (LangChain, Sourcegraph, Phil Schmid, Manus) and cited repeatedly by in-window work — which is why it is included despite the pre-window date.
Sub-agent context isolation in particular became a standard structural move, not just an optimization.
Caveat: out-of-window (September 2025) and a definitional position piece rather than an empirical study. Related token-savings figures cited elsewhere (Anthropic context editing and memory tool: 84% token savings, +39% on 100-turn tasks) are first-party and self-reported.

Comments

No comments yet.