Enhance Agent Reasoning and Prevent Failures

Ejentum Reasoning Harness is an MCP server exposing a library of 679 cognitive operations engineered in natural language, organized across four harnesses - reasoning, code, anti-deception, and memory - as MCP tools an agent calls when a task matches their trigger conditions. It targets four specific failure mechanisms in long agentic chains: attention decay (losing the original task), reasoning decay (compounding errors), sycophantic collapse (agreeing with the user's frame instead of evaluating it), and hallucination drift (asserting unsupported claims with confidence). Each harness call retrieves a task-matched scaffold rather than a fixed template, with five labeled fields the agent treats as internal-reasoning instructions, not output content: a named failure pattern (tagged [NEGATIVE GATE], [CODE FAILURE], [DECEPTION PATTERN], or [PERCEPTION FAILURE] depending on harness), a step-by-step [PROCEDURE] for an honest response, a [REASONING TOPOLOGY] control-flow graph, a [TARGET PATTERN] example of the corrected response shape, and a [FALSIFICATION TEST], or [VERIFICATION]/[INTEGRITY CHECK]/[PERCEPTION CHECK], applied post-draft. The agent ingests the scaffold and writes from it rather than from raw chain-of-thought, and its user-facing reply should stay in native voice with no echoed bracket names, procedural vocabulary, or meta-commentary about the harness. Invocation is on-demand - via automatic trigger matching or an explicit prompt like "Use harness_anti_deception, then answer: ..." - it does not auto-run every turn. Two worked examples: harness_anti_deception on a sunk-cost prompt about a three-month GraphQL gateway project separates past spending from prospective evaluation instead of anchoring on the investment already made; harness_code on a refactor where "all tests still pass" flags that phrase as a tool-shortcut signal rather than a correctness signal, surfaces call sites handling the old exception-raising behavior versus the new None-returning behavior, and recommends behavior-verifying tests before merge.

Use harness_reasoning before analytical, diagnostic, planning, or multi-step questions such as root-cause analysis, architecture decisions, or trade-off evaluation; harness_code before generating, refactoring, reviewing, or debugging code, or making architectural, algorithm, or dependency-upgrade decisions; harness_anti_deception when a prompt pressures the agent to validate, certify, or soften an honest assessment, or applies manufactured urgency or authority appeals; and harness_memory only to sharpen an observation already formed about cross-turn drift or behavioral patterns - never call it before an observation exists. Skip the harness entirely for simple factual lookups, syntax questions, file reads, code execution, or anything the agent can confidently complete in 1-2 steps natively. Its own best practices add explicit don'ts: never stack three or more harnesses in a single turn, since attention competition degrades the first call; never call harness_memory without an existing observation; and never treat the API as a hard dependency - on its typical 5-second timeout, fall back to native capability rather than blocking. Its own stated limitation: the harness shapes the substance of reasoning, it does not guarantee a correct answer, and domain expertise and source verification still apply - the scaffold is a reasoning procedure, not a knowledge base, so it retrieves no facts.

Input is the agent's in-progress task plus, for harness_memory, an already-formed observation to sharpen. Output is the five-field scaffold described above, which the agent absorbs internally before producing its actual, native-voice reply - the scaffold's bracketed field names and procedural language should never appear in that reply.

Installed as an npm-published MCP server via npx, added to a client's mcpServers config (Claude Code, Cursor, Cline, Windsurf, Codex CLI, Gemini CLI, Antigravity, or VS Code Copilot Chat) with an EJENTUM_API_KEY environment variable - a free tier offers 100 calls with no card required. Setup is manual; automated plugin installation is explicitly blocked for both the Codex and Claude plugin targets. The server makes outbound HTTPS requests only to the Ejentum Logic API gateway (Zuplo-hosted), authenticating via a Bearer token that must live in an environment variable or a client's secret-handling mechanism, never committed to source - it executes no shell commands and reads no filesystem paths beyond its own environment, making it a pure HTTP-proxy MCP server. Its own client config, quoted verbatim:

{
  "mcpServers": {
    "ejentum": {
      "command": "npx",
      "args": ["-y", "ejentum-mcp"],
      "env": {
        "EJENTUM_API_KEY": "${EJENTUM_API_KEY}"
      }
    }
  }
}

It is licensed MIT.

Agent builders and heavy AI-pair-programming users who want a structured, on-demand check against known agentic failure modes - losing the thread, compounding reasoning errors, caving to social pressure, or confidently asserting unsupported claims - applied selectively to the moments that actually need it rather than on every turn.