# Ejentum

> Reasoning Harness for Agentic AI. A REST API that injects engineered cognitive operations into LLM agents at inference time, blocking the four failure modes that cause multi-step agents to fail: attention drift, premature convergence, sycophantic acceptance of bad premises, and fabricated intermediate state.

Ejentum builds the **Reasoning Harness** layer for agentic AI. 679 engineered cognitive operations across 4 harnesses (reasoning, code, anti-deception, memory). Each operation is a structured procedure with a reasoning topology (executable DAG), failure patterns to block, correct reasoning procedure, verification tests, and suppression/amplification vectors. One API call. Works with any LLM. No fine-tuning.

Two delivery modes per harness:
- **Dynamic**: single retrieval, predictable injection. Cheaper, faster.
- **Adaptive**: top-k retrieval rewritten by an adapter LLM to the specific task identifiers. Cognitive operation reshapes itself for the problem at hand.

## Harnesses

- **Reasoning** (311 operations): 6 cognitive dimensions (causal, temporal, spatial, simulation, abstraction, metacognition).
- **Code** (128 operations): software-engineering layer — API grounding, guard migration, dependency management, language semantics.
- **Anti-Deception** (139 operations): 6 sub-layers — sycophancy, hallucination, deception, adversarial, judgment, executive.
- **Memory** (101 operations): perception layer — perception sharpening, behavioral calibration, state tracking, observation depth.

## API

Endpoint: `https://api.ejentum.com/harness/`

```
POST https://api.ejentum.com/harness/
Authorization: Bearer <api_key>
Content-Type: application/json

{"query": "your task description", "mode": "reasoning"}
```

Returns a pre-rendered cognitive injection (~400-900 tokens) to inject into the agent's context window before execution.

### 8 modes

Dynamic (single retrieval, all tiers):
- `reasoning`, `code`, `anti-deception`, `memory`

Adaptive (top-k retrieval rewritten for the task, Go/Super tier):
- `adaptive-reasoning`, `adaptive-code`, `adaptive-anti-deception`, `adaptive-memory`

Response key matches the mode name: `[{"reasoning": "..."}]`, `[{"adaptive-code": "..."}]`, etc.

### Injection structure

Six labeled blocks in canonical order:
1. **[PROCEDURE]** — natural-language reasoning steps
2. **[REASONING TOPOLOGY]** — executable DAG (Sn:, Gn{?}, M{...}, FREEFORM{...})
3. **[COGNITIVE PAYLOAD]** — Amplify: / Suppress: signals
4. **[FALSIFICATION TEST]** — verification criterion
5. **[NEGATIVE GATE]** — failure pattern to block
6. **[TARGET PATTERN]** — what correct reasoning looks like

Suppression signals (Suppress:) are the load-bearing element. They prune entire failure branches before generation.

## Pricing

- **Free trial**: 30 days, 1000 dynamic reasoning calls, no credit card.
- **Go**: €5/month. 1000 dynamic + 250 adaptive calls. All 8 modes.
- **Super**: €25/month. 5000 dynamic + 1500 adaptive calls. All 8 modes.

Sign up: https://ejentum.com/pricing

## Evidence

**Reasoning Harness:**
- EjBench (180 professional tasks): +10.1pp composite, +132% self-monitoring, +44% verification
- BBH / CausalBench / MuSR (110 academic tasks): 69.7% → 76.8% correctness, hardest tasks 20% → 60%
- ARC-AGI-3 (25 sequential game actions): 24-turn injection half-life, 12.2x reasoning depth

**Code Harness:**
- LiveCodeBench Hard (28 AtCoder tasks): 85.7% → 100% pass rate, zero regressions
- SciCode (10 hard scientific computing): zero bugs with dual injection

**Anti-Deception Harness:**
- ELEPHANT (40 sycophancy scenarios): 5.8% composite sycophancy rate, -5pp framing
- 20-turn adaptive adversarial attack: detected at Turn 6, 27/30 blind eval
- Hallucination prevention (5 fabrication tests): zero hallucinations

**Memory Harness:**
- State tracking (20-turn implicit changes): 50% fewer stale facts
- Perceptual detection (coaching conversations): 3x signal detection
- Blind evaluation: 4.1/5 vs 3.5/5

Open benchmarks: https://github.com/ejentum/benchmarks (CC BY 4.0)

## Why it works

Suppression (telling a model what NOT to do) is multiplicatively more effective than amplification (telling it what to do). A suppression signal prunes entire branches of the output space before generation. The reasoning topology gives the model a DAG to execute against, not a free-form chain that compounds errors. Validated across 250+ benchmark tasks.

## MCP server

The Ejentum Reasoning Harness is available as an MCP server. Install:

```
npx ejentum-mcp
```

8 tools exposed (4 dynamic + 4 adaptive). Works with Claude Code, Cursor, Zed, Continue, Cline, Goose, and any MCP-compatible agent runtime.

## Framework integrations

Native shims for: LangChain, LangGraph, CrewAI, Agno, AutoGen, Letta, LlamaIndex, smolagents, PydanticAI, Mastra, Vercel AI SDK, Genkit, n8n.

## Docs

- [Quickstart](https://ejentum.com/docs/quickstart): Add the reasoning harness to an LLM agent in 5 minutes.
- [API Reference](https://ejentum.com/docs/api_reference): Endpoints, request/response shape, error codes.
- [Method](https://ejentum.com/docs/method): The RA²R methodology behind the cognitive operations.
- [Reasoning Harness](https://ejentum.com/docs/reasoning_harness): 311 operations, 6 cognitive dimensions.
- [Code Harness](https://ejentum.com/docs/code_harness): 128 operations for code generation, refactoring, architecture.
- [Anti-Deception Harness](https://ejentum.com/docs/anti_deception): 139 operations blocking sycophancy, hallucination, prompt injection.
- [Memory Harness](https://ejentum.com/docs/memory_harness): 101 operations for perception, calibration, state tracking.

## Research

- [Why LLM Agents Fail](https://ejentum.com/blog/why-llm-agents-fail): Flagship essay on the four mechanism-level failures in production agents.
- [Under Pressure (Zenodo)](https://doi.org/10.5281/zenodo.19392715): Peer-reviewed paper on RA²R cognitive scaffolding. SSRN ID 6512038.

## Optional

- [Use Cases](https://ejentum.com/use-cases): 14 industry-specific failure patterns and resolution.
- [Abilities Registry](https://ejentum.com/abilities): Browse all 679 operations.
- [Benchmarks](https://ejentum.com/docs/benchmarks): Independent benchmark methodology and results.
- [About](https://ejentum.com/about): Founded in Athens, Greece, 2025.

## Company

Founded by Frank Brsrk in Athens, Greece. ORCID: 0009-0000-7086-6991. Contact: info@ejentum.com.