smolagents-ejentum

smolagents integration for the Ejentum Reasoning Harness. Exposes eight Tool subclasses (one per mode) plus an ejentum_tools() factory that returns all eight as a list.

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.

Four dynamic tools (reasoning, code, anti_deception, memory) are available on all tiers including the 30-day free trial. Four adaptive tools (adaptive_reasoning, adaptive_code, adaptive_anti_deception, adaptive_memory) additionally run an adapter LLM that rewrites the matched operation with task-specific identifiers; they require the Go or Super tier.

smolagents validates the Tool.name class attribute against a Python-identifier regex at __init_subclass__. Tool names here use underscores; the on-wire API mode strings (sent in the POST body) stay hyphenated. The split between LLM-facing name and on-wire mode lives in each tool's name vs mode class attributes.

Install

pip install smolagents-ejentum

Configuration

export EJENTUM_API_KEY="ej_..."

Or pass api_key= to any tool constructor. Get a key at ejentum.com/pricing.

Usage

All eight tools

from smolagents import CodeAgent, InferenceClientModel
from smolagents_ejentum import ejentum_tools

model = InferenceClientModel(model_id="meta-llama/Llama-3.3-70B-Instruct")
agent = CodeAgent(tools=ejentum_tools(), model=model)

agent.run(
    "We have spent three months on the GraphQL gateway. "
    "Should we keep going or pivot to REST?"
)

One tool

from smolagents import CodeAgent, InferenceClientModel
from smolagents_ejentum import EjentumAntiDeceptionTool

tool = EjentumAntiDeceptionTool()
agent = CodeAgent(tools=[tool], model=InferenceClientModel("..."))

Explicit API key

tools = ejentum_tools(api_key="ej_...")

Tool inventory

Dynamic (all tiers)

Adaptive (Go or Super tier)

Every tool defines name, description, inputs={"query": {...}}, output_type="string". Returns the injection from forward(query) as a string. Errors return as strings; forward does not raise.

API reference

# Per-tool (same constructor on every Ejentum*Tool class)
EjentumReasoningTool(
    api_key: str | None = None,
    api_url: str = "https://api.ejentum.com/harness/",
    timeout_seconds: float = 10.0,
)

# Factory
ejentum_tools(
    api_key: str | None = None,
    api_url: str = "https://api.ejentum.com/harness/",
    timeout_seconds: float = 10.0,
) -> list[Tool]

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 tools are hosted as an MCP server at https://api.ejentum.com/mcp. smolagents has a first-class MCP client (MCPClient, ToolCollection.from_mcp) that consumes the endpoint with Bearer auth.

Hub-shareable tools constraint

smolagents' Hub-share convention requires __init__ to take only self. The tools here are distributed via PyPI, not the Hub, so they take constructor args (api_key, api_url, timeout_seconds). If you fork and push_to_hub, refactor to hard-code config as class attributes first.

Compatibility

• Python 3.10+
• smolagents>=1.0.0
• 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.