Does this demo actually call the Jev API?

No. jev_classifier implements the same state plus noul, choice, and score contract locally with TF-IDF cosine similarity. The README names orchestrator/app/jev_client.py as the one-file swap point for the live model.

Why do the route instructions echo event wording?

The stand-in classifier is lexical rather than semantic, so state text and instruction wording must share vocabulary to discriminate. The README warns abstract descriptions score poorly and writes its route instructions to echo likely event words.

What runs without an Anthropic key?

Everything still runs end to end: LLM steps fall back to deterministic mock logic, and both pytest suites run standalone with mocked HTTP in well under a second, needing no containers, network, or keys.

How do I see the guardrail actually block something?

Post the flagged compliance sample and watch escalations.jsonl: the SHP-9931 case with incomplete docs and a customs flag is blocked from auto-approval, while the clean SHP-2077 case passes the guardrail.

What are the three orchestrator nodes?

A router that asks Jev a Choice question per event, a guardrail reached only for the compliance_review route that asks a Noul blocking question before auto_approve_shipment, and an executor that runs extract_fields, auto_approve_shipment, and notify_downstream.

How do I run the demo?

Copy .env.example to .env (optionally adding ANTHROPIC_API_KEY) and run docker compose up --build; the event-producer then posts the four sample events on a loop, and you can tail data/outbox.jsonl and data/escalations.jsonl.

How do I swap in the real Jev model?

It is a one-file change: jev_classifier implements the contract locally with a TF-IDF/cosine classifier, and the README names orchestrator/app/jev_client.py as the swap point for the live model.

CommunityJev Agents & MCP

Agent Harness With Jev LLM

Demo of the LangChain Jev harness pattern in LangGraph: a fast classifier routes events and gates a risky tool call, leaving open-ended reasoning to a chat LLM. Entry is POST /events; output lands in JSONL files.

See highlights

Open github.com

Stars: 1
Language: Python
Published by: Community
Author: dhirajpatra
Last verified: Oct 2026

dhirajpatra/agent-harness-with-jev-llm

User repository on GitHub

View on GitHub

A small multi-agent harness built around the pattern from LangChain's post: use a fast, non-generative "System One" classifier (TypeSafe AI's Jev) for structured in-loop decisions -- routing and risky-tool-call gating -- and reserve a real chat LLM for the genuinely open-ended reasoning.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A small multi-agent harness built around the pattern from LangChain's "Building a Harness with Jev" post: use a fast, non-generative "System One" classifier (TypeSafe AI's Jev) for structured in-loop decisions -- routing and risky-tool-call gating -- and reserve a real chat LLM for the genuinely open-ended reasoning.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Routing Jev for Tool calling Jev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Router asks Jev a Choice question per event and only falls back to the LLM below a confidence threshold.
Guardrail asks Jev a Noul block question before the compliance route may call auto_approve_shipment.
Local TF-IDF classifier implements the real Jev contract; swapping in the live model is a one-file change.
Docker Compose runs orchestrator, classifier, and event producer; pytest suites run with no keys or network.

Watch out

The bundled classifier is lexical TF-IDF similarity, not the trained Jev model, and LLM steps use deterministic mocks without a key. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

vexjoy-agent

441

A four-layer agent toolkit for Claude Code, Codex, Factory, and Reasonix: 43 agents, 62 skills, 78 hooks, 147 scripts, and a Jev-backed router that gates dispatch on intent alignment.

Jev Agents & MCP#community#jevbest

Community

notque/vexjoy-agentPython

muse-jev-playbook

Agent-oriented playbook for using Jev as a cheap, fast decision layer that triages, classifies, scores, and gates work before expensive browser, research, retry, or subagent steps.

Jev Agents & MCP#community#awesomejev#catalog-import

Community

Bodila51/muse-jev-playbookPython

Tiermem

Tiered agent memory keeping compact summaries with source links over searchable raw history, where a router decides per query whether notes suffice or raw pages must be recovered.

Jev Agents & MCP#community#jevbest#catalog-import

Community

FreedomIntelligence/TiermemPython

liteagents

Python SDK that runs the same prompt, tools, and MCP setup across DeepAgents, Pydantic AI, Claude Agent SDK, Codex, and OpenCode by switching harness.

Jev Agents & MCP#community#jevbest#catalog-import

Community

BerriAI/liteagentsPython

typesafe-skill-router

Opt-in Hermes plugin that routes each turn to one skill: a wide Jev Choice plus request checks gate a narrow fits round, injecting a single ignorable hint line.

Jev Agents & MCP#community#jevbest#catalog-import

Community

DECRUX9812/typesafe-skill-routerPython

tink-route

CLI that routes a task to one specialist skill from a Tink library using TypeSafe Jev, verifies the winner with tink mount, then prints it inline or as a mounted path.

Jev Agents & MCP#community#awesomejev#catalog-import

Community

jon-devlapaz/tink-routePython

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai