Context Engineering for 2026.
PH—010
Prompt engineering was never the practice. Context engineering is. How the 2026 Corpus, a live readiness score, and a per-industry ontology compose into context the agent actually uses. And what it costs, in tokens and in latency, to get this right.
I The Three Things Every Call Needs
Every agent call on our installs composes context from three sources. Not two. Not four. Three.
First, the base model’s general knowledge. We do not retrain it. We rely on it.
Second, the 2026 corpus slice. What the commercialization research says is currently shipping for the problem at hand. This is retrieved, not baked in.
Third, the installation’s own state. What the customer’s readiness score is, what the ontology filter is currently enforcing, what the last three agent calls produced. This is live, not cached.
The three sources answer three different questions. The base model answers “what is generally true.” The corpus slice answers “what is currently true.” The installation state answers “what is true for us.”
II Static vs. Rolling Context
Static context is the corpus slice and the ontology filter. It updates quarterly. It is cached, compressed, and injected on every call.
Rolling context is the installation state. It updates continuously. It is pulled fresh, but bounded. We never inject the last hundred agent calls. We inject the last three to five, plus the current readiness score, plus any unresolved escalations.
The discipline is that static context is bigger and rolling context is smaller. Not the reverse. Teams that let rolling context grow unbounded end up with context windows that reason primarily about their own recent output. That is a degenerate condition.
III What Context Costs
Tokens and latency, per call, across our installs:
- Base model context (system prompt, tool schemas): 1,200 – 2,400 tokens.
- Corpus slice: 3,000 – 6,000 tokens, compressed.
- Ontology filter directives: 400 – 900 tokens.
- Rolling installation state: 800 – 1,800 tokens.
Total per call, typical: 6,000 to 10,000 tokens before the user’s input is added. This is not cheap. It is also not optional. We have benchmarked installs against a thinner context and watched the ontology filter fail silently. The savings vanish the first time a non-compliant recommendation ships.
Latency: between eighty and two hundred milliseconds of context assembly overhead per call. Within budget for interactive use. Negligible for batch.
IV Context for Multi-Agent Hand-Offs
When one agent hands off to another, we do not pass the full context. We pass the structured output of the first agent and let the second agent rebuild its own context. The alternative (passing all three layers through) inflates the context of the downstream agent with material it does not need, and often makes the downstream agent reason about the upstream agent’s work rather than its own task.
The test for whether the hand-off is correct: if the downstream agent could complete its task given only the structured output, the hand-off is correct. If it needs upstream context to make sense of the task, the task was wrongly decomposed.