Context Engineering: The Discipline (2026)

Prompt engineering was 2023's craft. Context engineering is 2026's — and it is an engineering discipline, not a writing skill.

The definition

Context engineering is the systematic design of what tokens enter a model's context window, in what order, at what freshness, and at what cost — across the entire request lifecycle.

It replaces prompt engineering because in a 2026 agentic system, the user's message is <5% of what reaches the model. The other 95% — system prompt, tool schemas, retrieved chunks, memory, prior turns, sub-agent outputs — is the engineered surface.

The seven layers

1. Stable system prefix (cached)
2. Persona / examples / policies
3. Tool schemas (scoped per turn)
4. Retrieved evidence (RAG chunks)
5. Turn history (pruned / summarised)
6. Reasoning budget
7. Output shaping (schemas, length)

Each layer has its own cost model, cache behaviour, and update cadence. Optimising each in isolation misses interactions — hence the "engineering" label.

The token budget

Layer	Typical budget	Cache-friendly?
System prefix	2–4K	Yes (always cache)
Persona / policies	1–2K	Yes
Tool schemas	3–10K	Yes if scoped
RAG chunks	2–8K	No
Turn history	1–4K	Partial
Reasoning	0–16K	No (output-priced)
Output	0.5–2K	No

A feature exceeding its budget is a defect, not a request for a bigger budget. The fix is compression, not expansion.

The five techniques

1. Compression. LLMLingua-2, prompt-token-reduction (up to 20× on long documents with <2% quality loss).
2. Scoping. Load only the tool schemas relevant to the current turn (MCP scoping).
3. Summarisation checkpoints. After N turns, summarise history to 500 tokens.
4. Late binding. Inject dynamic content after the cached prefix, never before.
5. Reference over inclusion. Pass document IDs; let a tool fetch on demand.

The measurement

Add to every request span:

• context_tokens_by_layer (dict)
• context_utilization
• layer_cache_hit_ratio
• context_headroom

The 2026 stack

• LLMLingua-2 (Microsoft) — task-agnostic compression.
• dspy / TextGrad — automated prompt/example selection under a token budget.
• Anthropic contextual retrieval — chunk-level context prepending; +30–50% RAG accuracy.
• OpenAI Prompt Optimizer (2026 feature) — recompresses stable prefixes.

The role

2026 org charts increasingly include a Context Engineer — owning the prompt-and-schema registry, compression pipelines, cache-hit KPIs, and per-feature token budgets. Where this role does not exist, context grows unboundedly until a cost incident forces cleanup.