AGENTIC SYSTEM 1 · DUAL-BRAIN ARCHITECTURE · docs

03. State & Context Engineering

In the design-os-system-one architecture, State is the single source of contextual truth ingested by the model to evaluate questions. Unlike conventional LLMs that require verbose meta-prompts (“You are an intelligent AI assistant… Please return strict JSON…”), System 1 ingests state as pure, unadorned data.

---

1. Supported State Formats

The System 1 engine supports heterogeneous data representations:

A. Plain Text Strings

Optimal for emails, customer support tickets, chat logs, code excerpts, and legal clauses.

{
  "state": "Hello, I have been unable to log into my account since enabling SMS 2FA this morning. The verification codes never arrive.",
  "questions": { ... }
}

B. Structured Objects (JSON Dictionaries)

Optimal when aggregating database entities, account metadata, and session telemetry.

{
  "state": {
    "user_tier": "Enterprise",
    "account_age_days": 450,
    "last_payment_status": "succeeded",
    "open_tickets_count": 3,
    "message": "Webhook delivery is currently delayed by over 15 minutes. We are dropping critical customer transactions."
  },
  "questions": { ... }
}

The model parses JSON keys and values to capture correlations between user tier, historical reliability, and incident severity.

C. Arrays of Candidate Items

Optimal for re-ranking, candidate filtering, or extraction:

{
  "state": {
    "query": "lightweight laptop for software engineering",
    "candidates": [
      {"id": "p1", "name": "ThinkPad X1 Carbon Gen 11", "weight": "1.12kg", "cpu": "i7-1365U"},
      {"id": "p2", "name": "ASUS ROG Strix G16", "weight": "2.5kg", "cpu": "i9-13980HX"},
      {"id": "p3", "name": "MacBook Air 15 M2", "weight": "1.51kg", "cpu": "Apple M2"}
    ]
  },
  "questions": { ... }
}

---

2. Golden Rules for State Hygiene

1. Pure Facts, Zero Meta-Instructions

❌ Anti-pattern: Embedding prompts and reasoning instructions inside the State:

{
  "state": "Below is a user message. Carefully analyze if they are angry: 'I want to cancel my subscription'."
}

✅ Correct: Cleanly isolate evidence into state and intent into instructions:

{
  "state": "I want to cancel my subscription.",
  "questions": {
    "is_cancellation": {
      "type": "noul",
      "instructions": "The customer requests subscription termination"
    }
  }
}

2. Context Pruning

Feeding raw HTML boilerplate, CSS class tokens, tracking scripts, and irrelevant database UUIDs degrades calibration and inflates processing latency.

• Strip HTML markup and provide clean Markdown or structured JSON.
• Prune extraneous fields; transmit only data relevant to the decision gate.

3. Semantic Key Labeling

When passing JSON objects, use self-documenting keys (customer_history, order_status, error_trace). System 1 leverages key semantics to locate evidence relevant to the question.

---

3. Context Budgeting & Long Documents

• Optimal Operating Range: System 1 decisions achieve highest precision and lowest latency when state is sized between 50 and 4,000 tokens (1–10 pages of clean text).
• Processing Extended Documents (Contracts, System Logs):
  • For massive corpora (> 8,000 tokens), employ BM25 pre-filtering or semantic chunking first.
  • Employ System 1 as the high-speed re-ranking or verification layer across chunks (refer to the Cookbooks section).