Context Engineering Is the New Prompt Engineering

paicontext-engineeringmiesslerarchitecture

Trimmed to bullet-point summary 2026-05-09 — original prose archived at _archive/blog-articles-pre-trim-2026-05-09/context-engineering-is-the-new-prompt-engineering.md. Pending rewrite in voice.

Thesis

  • Prompt engineering circa 2023 was a hack against stateless models with nothing to work with.
  • The actual leverage was always in the context pipeline behind the prompt.

The Shift

  • Prompt = thin API call. Everything interesting happens before the call: retrieval, memory, routing, world model the model is pointed at.
  • Prompt engineering: “How do I phrase this so the model responds well?”
  • Context engineering: “What does the model need to know about this entity, at this moment, for this decision — and how do I pipe it in reliably?”
  • One is wordsmithing; the other is systems design.

Miessler’s Thesis Progression

  • Same claim, sharpening over time:
    • 2023: SPQA (State, Policy, Questions, Actions)
    • mid-2025: “The Future of Hacking is Context”
    • Now: Unified Entity Context (UEC)
  • Core claim: software verticals dissolve; entity remains; AI sits on top of the entity model and acts.
  • Dinner vignette (Miessler): drunk employee at target company posts default-creds subdomain; agent pings; user authorizes probe; exploit confirmed by appetizers. Pipeline-driven, not prompt-driven.
  • Anthropic 2025 threat report: one actor hit 17 organizations in weeks via context pipeline, not jailbreak.
  • Attack/defense identical shape: whoever has cleaner, fresher context wins. Context speed is the new metric.

Author’s Stack: DAI

  • DAI = context pipeline with an LLM glued to the top. The model is not the product.
  • Entity model. Telos at center — explicit versioned representation of self (goals, beliefs, projects, problems, relationships, budget, health). Single source of truth. Telos changes first; memory and retrieval propagate.
  • Memory. Session continuity in memory/. Structured, references docs/, never contradicts. Long-term facts in docs/; short-term state in memory/; rendered views downstream. Top-down propagation.
  • Retrieval and routing. Lightweight router decides what context slice to load per request. Drowning the model in a 4M-token soup is anti-pattern.
  • Lifecycle hooks. 39 hooks fire on events. Hooks read source-of-truth, never write back. One writer, many readers.
  • The algorithm. Seven-phase loop (observe, think, plan, build, execute, verify, learn). The phase is where context gets assembled, pressure-tested, verified. Per-step prompts trivial.

Why Prompts Became Thin

  • Once entity context + memory + retrieval + routing do real work, the prompt shrinks to nothing.
  • 2023 version of “should I take this job?” — dense prompt with role-playing, constraints, three pasted documents.
  • 2026 version — one-line question. The pipeline already knows goals, projects, monthly burn, open offers, partner constraints, the last five conversations on similar moves.

Build Implications

  • Stop optimizing prompts in isolation.
  • Treat context as a product. Version it. Validate it. Monitor for drift.
  • Sarit Tager (Palo Alto): unified data lake across code/cloud/runtime/identity. Coverage before risk.
  • Enforce one source of truth per fact. Hierarchy is what keeps a living system coherent.
  • Watch the freshness gap (state change → system understands). Security AND productivity metric.
  • Bigger context window ≠ better system. 4M tokens is storage, not strategy.

Conclusion

  • Prompt engineering was a stepping stone — what we tell the model.
  • Context engineering = what the model can know, structurally, before we say a word.
  • Treat context as infrastructure: entity-first, hierarchical, versioned, always fresh. Pipelines over phrases.