crewai-ejentum

CrewAI tool for the Ejentum Reasoning Harness. Exposes a single EjentumHarnessTool class with a mode parameter covering eight values: four dynamic (reasoning, code, anti-deception, memory) and four adaptive (adaptive-reasoning, adaptive-code, adaptive-anti-deception, adaptive-memory).

Use the harness before the agent generates on complex, multi-step, or multi-constraint tasks where the model's default reasoning template would miss a constraint, take a shortcut, or drift across turns. Each call returns a cognitive operation: a structured procedure (numbered steps with a failure pattern to refuse and a falsification test) paired with an executable reasoning topology (a DAG of those steps with decision gates, parallel branches, bounded loops, and meta-cognitive exit nodes). The agent reads both layers before producing its response.

Dynamic modes return the top-1 abstract operation from the matching library; adaptive modes additionally run an adapter LLM that rewrites the operation with task-specific identifiers. Adaptive modes require the Go or Super tier.

Install

pip install crewai-ejentum

Configuration

export EJENTUM_API_KEY="ej_..."

EJENTUM_API_KEY is read from the environment at call time. Get a key at ejentum.com/pricing.

Usage

from crewai import Agent, Task, Crew
from crewai_ejentum import EjentumHarnessTool

harness = EjentumHarnessTool()

architect = Agent(
    role="Senior architect",
    goal="Evaluate technical decisions honestly",
    backstory="Pragmatic; pushes back on sunk-cost framings.",
    tools=[harness],
)

task = Task(
    description=(
        "We have spent three months on the GraphQL gateway. It's mostly done. "
        "Should we keep going or pivot to REST? "
        "Call the Ejentum harness with mode='anti-deception' first."
    ),
    agent=architect,
    expected_output="A recommendation that separates past spending from prospective evaluation.",
)

Crew(agents=[architect], tasks=[task]).kickoff()

Modes

Mode	Library size	Domain
reasoning	311	abstraction, time, causality, simulation, spatial, metacognition
code	128	software-engineering layer
anti-deception	139	sycophancy, hallucination, deception, adversarial framing, judgment, executive control
memory	101	perception layer (filter-oriented; not for fact extraction)
adaptive-reasoning	311 (same pool)	with adapter LLM rewriting procedure + topology for the specific task
adaptive-code	128	same as above for code
adaptive-anti-deception	139	same as above for anti-deception
adaptive-memory	101	same as above for memory

Inputs

EjentumHarnessTool._run accepts:

• query (string, required): a 1-2 sentence description of the task. For mode="memory" or "adaptive-memory", use the format "I noticed X. This might mean Y. Sharpen: Z.".
• mode (string, required): one of the eight mode strings above.

Returns the injection as a string. Errors return as human-readable strings; the tool does not raise, so an agent step never crashes the run.

API reference

EjentumHarnessTool(
    api_url: str = "https://api.ejentum.com/harness/",
    timeout_seconds: float = 10.0,
)

Field	Default	Description
api_url	https://api.ejentum.com/harness/	Override for self-hosted gateway.
timeout_seconds	10.0	Per-call HTTP timeout.

Wire contract

POST https://api.ejentum.com/harness/
Headers: Authorization: Bearer <key>, Content-Type: application/json
Body:    { "query": <string>, "mode": <one of 8 mode strings> }
Response (200): [ { "<mode>": "<injection string>" } ]
Response (401|403|429): { "error": "..." }

Full wire contract, field structure of an injection, DAG syntax, and a canonical dynamic-vs-adaptive comparison on the same query are documented in the ejentum-mcp README.

ejentum-mcp alternative

The same eight modes are exposed as MCP tools at https://api.ejentum.com/mcp. If you prefer that route, CrewAI's MCP support can consume the hosted endpoint with Bearer auth.

Compatibility

• Python 3.10+
• crewai>=0.40.0
• requests>=2.31.0

License

MIT

Measured effects

The Ejentum harness is benchmarked publicly under CC BY 4.0 at github.com/ejentum/benchmarks:

• ELEPHANT sycophancy: 5.8% composite on GPT-4o (40 real Reddit scenarios)
• LiveCodeBench Hard: 85.7% to 100% on Claude Opus (28 competitive programming tasks)
• Memory retention: 50% fewer stale facts served (20-turn implicit state changes)
• Plus per-harness numbers across BBH/CausalBench/MuSR, ARC-AGI-3, SciCode, and perception tasks

Methodology, scenarios, run scripts, and raw outputs are all in-repo.