Ejentum

Ejentum is a reasoning harness for agentic AI, built by Franko Luci (Frank Brsrk) and founded in March 2026. It operates as a tool your agent calls mid-loop, returning a structured cognitive operation matched to the task at hand — without retraining or prompt rewrites. The product targets a specific class of production failure: reasoning errors that compound silently across multi-step agent chains.

What It Is

Ejentum is an inference-time reasoning correction layer delivered as a REST API and hosted MCP server. Rather than baking reasoning strategies at build time, it retrieves one of 679 engineered cognitive abilities at runtime and injects it into the agent's context before execution. The injection includes a failure guard (the pattern to avoid), a step-by-step reasoning approach, a reasoning topology graph, suppression signals for known shortcuts, amplification signals, and a falsification check. The harness does not add new model capability — according to the product page, it removes the failure modes that were consuming existing capability.

Four Cognitive Harnesses

Ejentum organizes its 679 abilities across four product layers:

• Reasoning — Channels analytical power across six cognitive dimensions: causality, time, space, simulation, abstraction, and metacognition. Blocks shortcuts that flatten careful analysis into surface-level pattern matching.
• Code — Enforces correctness checks before the model commits to an approach, adds verification loops, and preserves safety guards that typically vanish during refactors.
• Anti-Deception — Suppresses flattery, fabrication, sunk-cost validation, and the pull to tell users what they want to hear rather than what is true.
• Memory — Tracks people, signals, and context drift across long conversations so the agent stops treating turn fifty like turn one.

Each harness supports a dynamic mode (returns the best-matching ability as engineered) and an adaptive mode (rewrites the cognitive operation to fit the specific shape of the task).

Benchmark Evidence

Ejentum publishes benchmark methodology, raw data, and negative findings in a public GitHub repository. The vendor-published results across eight evaluation suites include:

• Code (LiveCodeBench Hard, 28 tasks, Claude Opus 4.6): Pass rate moved from 85.7% to 100%, with zero regressions.
• Code (SciCode, 10 tasks, Claude Opus 4.6): Scientific computing bugs dropped from 7 to 0; blind evaluation chose the injected output 10/10 times.
• Reasoning (ARC-AGI-3, 25 steps): Reasoning depth trend increased 12.2x; injection persisted 24 steps; memory decay reversed.
• Anti-Deception (ELEPHANT, 40 Reddit scenarios, GPT-4o): Composite sycophancy rate measured at 5.8%.
• Memory (perceptual detection, 15-turn scenario, GPT-4o): Signal detection rate tripled (14% to 43%).

The benchmark repository also reports negative findings: correctness dipped under reasoning-multi on EjBench, spatial domain regressed under reasoning-multi on BBH, and ARC-AGI-3 Level 0 was not cleared by either condition. The research paper "Under Pressure: RA²R and the Emergence of Uninstructed Reasoning Behaviors in Scaffold-Augmented Language Models" (Franko Luci, April 2026) is published on Zenodo and SSRN.

Integration and Deployment

Ejentum connects via a single POST https://api.ejentum.com/harness/ endpoint (Bearer auth, JSON in, JSON out) or a hosted MCP server at api.ejentum.com/mcp. The product page lists native support for:

• Frameworks: LangChain, LangGraph, CrewAI, LlamaIndex, Pydantic-AI, Agno, AutoGen, Smolagents, Mastra, Flowise, Langflow, Botpress, Voiceflow
• No-code platforms: n8n, Make.com, Zapier, Heym
• IDEs and coding agents: Cursor, Windsurf, Claude Code, Codex
• Model providers: OpenAI, Anthropic, Google, Meta, Mistral, Groq, Cohere, Hugging Face, Amazon Bedrock, Microsoft Azure, xAI, DeepSeek, Fireworks AI, Nous Research, Perplexity, Replicate, Inception Labs

The abilities are described as model-agnostic structured reasoning injections that work with any instruction-following model.

Honest Scope and Tradeoffs

The founder explicitly states on the product page that Ejentum does not help every agent: single-step classifiers, simple RAG lookups, and tasks where the model already converges in one hop are not good fits. The harness is positioned for multi-step chains where errors compound — planning agents, research agents, and code agents touching more than a handful of files. The FAQ notes that when the adaptive call pool is exhausted, calls silently fall back to their dynamic equivalent with a header flagging the fallback, so the agent continues working.