A library of multi-agent teams (agentic workflows) using Ejentum's RA²R Ejentum API as the cognitive layer. Each team is a deployable artifact: drop it into your runtime, configure one credential, run.
The teams in this repo are tuned for specific tasks where multi-cognitive analysis genuinely beats single-agent output. They use cross-lab agent diversity (Anthropic, Google, Alibaba, Zhipu, OpenAI) to reduce correlated failure modes that come from any single model family.
A 4-agent team (architect + reasoner + implementer + reviewer) for heym v0.0.9+ that performs deep adversarial code review. Refuses to rubber-stamp PRs, sources every concern from a specialist agent, grounds approvals in specific positive evidence. Each specialist applies a distinct Ejentum cognitive harness (reasoning, code, anti-deception). → Setup and verification tests
Evaluation framework for comparing baseline LLM output against Ejentum-augmented LLM output, with a third-party blind judge. Available as n8n workflows, a Python CLI, and integration patterns for agentic IDEs (Cursor, Antigravity, Claude Code). The instrument is the artifact: import, run on your own prompts, see the diff. → Spec and integrations
A single-HTML side-by-side blind evaluation module for any OpenAI-compatible LLM, with optional Ejentum harness wired in as a tool call. Two agents answer the same prompt; a blind judge scores both; results revealed only after scoring. One stdlib Python proxy serves the HTML, no build step, no framework, no install. → Setup and how to wire it
A 3-agent team for heym v0.0.20+ that performs pre-commitment self-evaluation for agent runtimes. Submit (task, planned method); three blind cross-lab evaluators (steelman defends, stress_test attacks, gap_finder finds what's missing) return three independent perspectives. Calling agent integrates them — no synthesizer. Models cannot reliably self-evaluate; this is the structural fix. Validated across engineering refactors, payments migration, security incident response, investigative reasoning. → Setup and verification tests
A 4-agent team for heym v0.0.30+ that turns a raw blood panel into a structured patient-education report. Built around a deterministic Python safety gate: a 12-marker hospital panic-value table runs synchronously inside the orchestrator BEFORE any LLM reasoning, and short-circuits to a fixed EMERGENCY OUTPUT on panic values without calling any sub-agent. Non-emergency panels fan out to three role-locked specialists in parallel: a plain-language interpreter (qwen3-max-thinking + reasoning), an anti-deception second-opinion voice (claude-opus-4 + anti-deception), and a differential enumerator (deepseek-r1, no harness). Cross-lab model diversity + per-agent Ejentum cognitive harnesses attached via MCP streamable_http. Patient-education only, NOT a diagnostic tool. → Setup and verification tests
The smallest working setup for the Ejentum Reasoning Harness in n8n: one AI Agent with the four n8n-nodes-ejentum community-node operations (reasoning, code, anti-deception, memory) wired onto its Tools input. No HTTP nodes, no header-auth wiring. Install the community node, set two credentials, import, and the agent autonomously picks a harness per task. Start here, then move to the four-pattern template below. → Setup and how it works
Four ways to wire a reasoning harness into an n8n agent, each with a different control vs. flexibility tradeoff. One importable workflow, one chat trigger, four branches selected by prefix: dynamic system prompt (you pick the mode), reasoner agent (one tool, model decides when), full harness (four tools, model decides which), and ejentum-mcp (single MCP node, same as full harness with smaller footprint). Pick your tradeoff. → Setup and the four patterns in detail
Each agent (or evaluation step) calls the Ejentum API in one of four cognitive modes:
| Mode | Purpose |
|---|---|
reasoning |
General reasoning scaffold — failure-mode suppressors, target patterns, falsification tests |
code |
Code-specific failure suppressors — cross-layer mismatch, environment drift, invariant violations |
anti-deception |
Suppresses sycophancy, hallucination, prompt injection, false certainty |
memory |
Perception sharpening, behavioral calibration, cross-turn observation |
Two paths to wire each cognitive mode into a specialist agent:
- HTTP Request tool (canonical): the curl pattern shown in each team's README. Works on every runtime.
- MCP server (when the runtime supports MCP clients): use ejentum-mcp. Two install paths: stdio via
npx -y ejentum-mcpfor clients that spawn MCP servers as subprocesses (Claude Desktop, Cursor, Windsurf, Codex CLI, Claude Code), or hosted HTTPS athttps://api.ejentum.com/mcpfor HTTP-MCP clients (n8n MCP Client and others). Either way the eight tools (four dynamic plus four adaptive) appear with their canonical names your specialist can call directly. No HTTP wiring per agent.
Get an Ejentum API key at ejentum.com/pricing. 30-day free trial; adaptive tools require Go or Super tier.
Read more at ejentum.com/docs.
The pattern (orchestrator + N specialists, each with a mode-specific harness) generalizes. To contribute or fork a new team:
- Pick a target task with multi-cognitive value (refactor planning, security audit triage, production debug forensic, spec verification, test design from traces, etc.).
- Decide which Ejentum modes each specialist needs.
- Build the workflow on your runtime of choice (heym, n8n, LangGraph, your own).
- Drop into a new folder here with:
README.md, workflow file(s), skill MDs (if applicable), system prompts, screenshots. - Add a row to the Catalog above.
Future teams in the planning phase: refactor planner, security audit triage, production debug forensic.
MIT. See LICENSE.