Architecture

How the harness API works, what it returns, and how your agent uses it. For the theoretical foundations, see The Method. For the endpoint spec, see API Reference.

Request Lifecycle

Every call to /harness/ passes through four stages. In dynamic mode, the same input always produces the same output at zero LLM inference cost; adaptive mode adds one adapter-model pass that rewrites the matched operation to your task.

POST /harness/
  |
  1. Authentication     Verify API key, resolve tier, enforce rate limits
  2. Retrieval          Match query against 679 abilities across 4 harnesses
  3. Composition        Assemble the cognitive injection (adaptive rewrites it to your task)
  4. Response           Return pre-rendered injection string as JSON

Dynamic retrieval executes in under one second with no LLM call. The response is a structured injection that channels the model's existing power into disciplined execution.

What the API Returns

Dynamic Modes (reasoning, code, anti-deception, memory)

One cognitive operation: the highest-scoring match for your query from the selected harness, returned as-is.

[
  {
    "reasoning": "[PROCEDURE]\nStep 1: Trace backward from failure...\n\n[REASONING TOPOLOGY]\nS1:identify_failure -> S2:trace_backward...\n\n[COGNITIVE PAYLOAD]\nAmplify: reverse replay; counterfactual node test\nSuppress: writing a vague post mortem summary...\nCognitive Style: backward chain analysis\nElasticity: coherence=root cause isolation, expansion=adversarial\n\n[FALSIFICATION TEST]\nIf an error's origin is not traced by replaying...\n\n[NEGATIVE GATE]\nThe checkout service crashed and we documented the incident...\n\n[TARGET PATTERN]\nReverse replay from the crash point..."
  }
]

Adaptive Modes (adaptive-reasoning, adaptive-code, adaptive-anti-deception, adaptive-memory)

The same matched operation, but an adapter model rewrites the [PROCEDURE] and topology to name your task's specifics. Every other section is returned identical to the dynamic response.

[
  {
    "adaptive-reasoning": "[PROCEDURE]\n...rewritten with your task's variables...\n\n[REASONING TOPOLOGY]\n...rewritten with your task's nodes...\n\n[COGNITIVE PAYLOAD]\nAmplify: ...\nSuppress: ...\n\n[FALSIFICATION TEST]\n...identical to dynamic...\n\n[NEGATIVE GATE]\n...identical to dynamic...\n\n[TARGET PATTERN]\n...identical to dynamic..."
  }
]

Response Components

Six labeled blocks in canonical order. Labels differ per harness; the function of each slot is the same.

Labels differ per harness: reasoning uses [NEGATIVE GATE], code uses [CODE FAILURE], anti-deception uses [DECEPTION PATTERN], memory uses [PERCEPTION FAILURE]. Each harness also returns a correct-shape block ([TARGET PATTERN] / [CORRECT PATTERN] / [HONEST BEHAVIOR] / [CLEAR SIGNAL]) and a [COGNITIVE PAYLOAD] block carrying the Amplify / Suppress / Cognitive Style / Elasticity lines. Adaptive modes return the same sections; only the procedure and topology are rewritten.

Product Layers

Each mode routes to a specialized ability collection. The query is matched against the abilities within that layer.

A supply chain risk query scores high on Causal and Temporal. A market expansion question scores on Spatial, Simulation, and Abstraction. The routing is probabilistic, not categorical.

How to Inject

Wrap the API response in delimiters and prepend to your agent's system message, BEFORE the task prompt:

[REASONING CONTEXT]
{paste the response value here (key matches the mode name)}
[END REASONING CONTEXT]

Now complete the following task:
{your agent's actual task}

The injection must come BEFORE the task. The suppression vectors need to be in context when the agent begins reasoning, not after.

For multi-turn agents: re-inject per turn. Each turn may activate a different reasoning dimension. Stale injections from previous turns degrade as context fills with task-specific tokens.

Token Overhead

Compare to a typical system prompt: 5,000 to 15,000 tokens. The injection is compact by design. It replaces prose with structured constraints.

Integration

Your agent calls Ejentum like any other tool. One POST request, one JSON response, one injection.

Request

POST https://api.ejentum.com/harness/
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json

{
  "query": "Why did our deployment fail after the config change?",
  "mode": "reasoning"
}

Compatible Frameworks

Ejentum works with any system that can make an HTTP POST and inject text into a prompt:

• n8n. AI Agent node with HTTP Request Tool
• LangChain / LangGraph. Custom tool or LCEL chain
• CrewAI. Tool injection into agent backstory
• Claude Code / Agent SDK. tool_use definition
• Cursor, Windsurf, Antigravity, Codex. custom HTTP tool
• Any HTTP client. Direct POST, parse JSON, inject

See the Integrations guide for framework-specific code examples.

Rate Limits and Errors

For the full API specification, see the API Reference.