Does it claim LLM-free execution or general cost savings?

No. The loop diagrams state that the shown iterations used LLM authoring, and the paired runs are single historical turns with separate workspaces. The built-in paired benchmark exists to make authored-heavy versus navigation-heavy tradeoffs visible.

How do Jev and the LLM avoid disagreeing on state?

The durable Transcript is the single source. The LLM reads a conversation projection and Jev reads a bounded Workspace with grounded candidates; both views are rebuilt from the ledger with independent display budgets.

How do parallel typed tool decisions stay valid?

Mounted ToolSpec declarations compile into one Jev request with an operation question plus speculative target questions, and only the target head compatible with the chosen operation is consumed. Invalid distributions fail before execution.

What is the sandbox and safety story for tool effects?

Containers run as non-root with dropped capabilities, read-only root filesystem, and resource limits. External writes are fail-closed behind confidence policy, recipient allowlists, budgets, and durably recorded intents before dispatch.

What language and license is JevLoop, and is it stable?

Python 3.12 or newer, Apache-2.0 licensed, and marked pre-alpha. It is an independent open-source project powered by Jev, not an official TypeSafe product.

How do I run its paired benchmark?

Run make bench, or cd backend and run uv run jevloop bench. Reports land in backend/artifacts/bench/ /report.{json,md} and record the digest, model identities, checked expectations, routing split, cache-aware cost, wall-clock time, and observed advantage.

What execution profiles does the runtime offer?

single_shadow runs JevLoop with executed sandbox effects and dry-run ledger writes; single_live executes the writes after guardrails; paired_shadow runs JevLoop against a plain LLM lane in isolated workspaces with dry-run writes in both. Paired live writes are intentionally not representable.

CommunityJev Practices & Patterns

Jevloop — parkavenue9639

Independent open-source agent runtime where Jev makes typed tool decisions, an LLM authors text only when needed and reviews low-confidence calls, and code owns execution.

See highlights

Open github.com

Stars: 7
Language: Python
License: APACHE-2.0
Published by: Community
Author: parkavenue9639
Last verified: Oct 2026

parkavenue9639/jevloop

User repository on GitHub

View on GitHub

A Jev-driven general-purpose agent harness for faster, lower-cost execution, with built-in side-by-side experiments against LLM-only agents.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is JevLoop — zjunlp.

About this project

JevLoop is an open-source agent runtime powered by TypeSafe Jev. Jev handles fast, calibrated tool decisions; an LLM authors language only when needed and reviews low-confidence choices; deterministic code owns execution, safety, durability, and state. The result is a conventional tool-using agent with a fast path: bounded choices avoid full LLM calls, while authored content still comes from an LLM with the complete conversation transcript and provider prompt-cache continuity.

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

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Moderation Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Jev decides operations and targets through validated probability distributions; the LLM writes text and reviews low-confidence calls.
One ledger stores conversation and execution evidence; Jev and LLM projections rebuild from it under a projection contract.
Observations become grounded argument candidates, or LLM_PARAMETERS when the same operation needs authored arguments.
paired_shadow runs JevLoop beside a plain LLM loop with separate transcripts and sandboxes for direct comparison.

Watch out

Pre-alpha and independent of TypeSafe; paired-benchmark figures describe one recorded run each, not a general speed or cost claim. 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

jev-guardrails

Mock carrier support agent with the guardrail layer built twice behind one interface: chat-model JSON judging versus TypeSafe Jev typed questions over the same 25 rules.