Context Engineering: The Orchestration Layer That Decides What the Model Sees
Prompting tunes words; context engineering tunes the model's evidence.
Image backend outage: the primary visual renderer returned OpenRouter HTTP 402 (insufficient credits). This mobile-first text and CSS card preserves the teaching payload; no PNG was generated and no alternate image backend was used.
The Window Is a Budget
- Nine sources compete for one finite token window.
- Pin the system contract and user context.
- Select retrieval, history, tools, data, memory, constraints, examples.
- Choose what to admit, order, compress, or drop.
- Prompt tuning changes one slot; orchestration changes the whole window.
The Worked Trace
- An invoice question needs fresh billing-tool JSON.
- It also needs the account id from memory.
- Add the refund policy from retrieval.
- Old small talk is not required evidence.
- Reserve headroom before the tool result arrives.
Exclusion Is the Design
- Admitting all nine sources every turn is the failure mode.
- More context helps, peaks, then hurts.
- Long windows lose the middle; distractors lower accuracy.
- Position changes behavior, so order is load-bearing.
- Feedback needs a judge that scores context use.
Falsify Context Quality
- Ablate retrieval, memory, and examples one at a time.
- Halve history on a fixed evaluation set.
- Then inject one irrelevant but plausible passage.
- A good system's score should drop.
- If it does not, the evaluation cannot see context quality.