Context Engineering: The Real Moat in the Agent Era

2025 was prompt engineering. Top teams in 2026 optimize context engineering—the prompt is a small slice of what the model sees.

From Prompt to Context

Prompt Engineering (2023–2024):
  Focus → better instructions
  Scope → user's last message

Context Engineering (2025–2026):
  Focus → full information environment
  Scope → system prompt + skills + retrieved docs + tool results
         + history + files + logs + …

Same prompt, different context assembly → 10× quality swing (Anthropic, Cursor engineering blogs).

What Goes in Context

One agent call might include:

┌─ System prompt ─────────────────────────┐  ~500 tokens
├─ Skills (on demand) ───────────────────┤  ~1000 tokens
├─ MCP resources ────────────────────────┤  ~2000 tokens
├─ Code context ─────────────────────────┤  ~8000 tokens
├─ Tool results ─────────────────────────┤  ~3000 tokens
├─ Conversation history ─────────────────┤  ~5000 tokens
├─ User message ─────────────────────────┤  ~200 tokens
└────────────────────────────────────────┘
                              ~20,000 of 128,000

Still problems:

1. Lost in the middle—weaker attention mid-context
2. Noise—irrelevant files steer the model wrong
3. Linear cost—every +1000 tokens adds latency and spend

Six Context Strategies

1. Retrieve Precisely, Don’t Dump Everything

❌ Entire 500-file repo in context
✅ Semantic index → load 5–10 relevant files

Cursor @file = precise; @codebase = search then select.

2. Layered Summarization

Raw chat (5000 tokens)
  → summary (500): "User refactoring auth; JWT done; writing tests"
  → keep summary + last 3 full turns

Claude Code compresses early steps on long tasks.

3. Selective Tool Results

Ten tool calls—not all ten stay:

ls → keep (structure)
cat config.json → keep
npm test → keep failures only (2000 lines → 20)
debug steps 4–7 → drop (superseded)
final fix → keep

4. Dynamic Skill Loading

Load 1–3 skills for “create PR”—not all 20 skills every turn.

5. Structured Beats Prose

❌ "TypeScript project, Astro 5, Tailwind 4, Cloudflare Pages, no DB…" (~100 tokens)

✅ stack: { lang: "TS", framework: "Astro 5", css: "Tailwind 4",
           deploy: "CF Pages", db: null } (~30 tokens)

6. Position Matters

[System prompt]     ← rules (always seen)
[… bulk middle …]
[User message]      ← current task
[Latest tool result] ← fresh evidence adjacent to user

Cursor’s Context Stack (Observed)

1. Built-in agent system prompt
2. .cursor/rules/
3. Matched SKILL.md files
4. Codebase semantic index (top-K files)
5. Open editor tabs (higher weight)
6. Recent terminal output
7. MCP resource payloads
8. Session chat history

Index retrieval quality = whether the agent finds the right files.

Claude Code on Long Tasks

• Start: relevant files + CLAUDE.md
• Mid-run: compress completed work to summaries
• Errors: keep message + stack trace, not full logs
• Late steps: “current state + remaining work + recent actions”

Enables 50+ step runs without forgetting the original goal.

Measuring Context Quality

Metric	Meaning	How
Context precision	% of context that’s relevant	human labels
Context recall	needed info included?	vs ideal context
Token efficiency	tokens for same quality	A/B strategies
Needle-in-haystack	find buried fact?	benchmarks

Summary

Prompt = what you say. Context = what the model sees. Optimizing context beats tweaking prompts by an order of magnitude.

Highest ROI inside the harness. Next: measuring whether agents actually improve—eval and observability.