Is this a finished autonomous agent I can deploy?

No. The core profile runs with no weights or key as a deterministic protocol test, the CLI needs explicit plan plus approve for side effects, and the 159 passing tests cover runtime and protocol rather than agent quality.

How can 255 options cover hundreds of tools?

Through virtual pages with stable IDs and revisions: PAGE widens candidate coverage while REFINE deepens one resident option, staying within the 255-option request ceiling.

What stops an uncertain file write from retrying silently?

It does not. An unknown receipt puts the task in needs_reconciliation until an external observer confirms the result, and validation runs through verdict and receipt protocols.

What do the 159 passing tests actually prove?

That the paging, context, memory, approval and recovery mechanics hold under scripted workloads. Live paging and recovery reports mark failure boundaries rather than claiming production readiness.

How does it keep the context window bounded?

With bounded context residency built on pinned blocks, aging, revisions, dependency invalidation and stable block IDs. Material can be refreshed or replaced in place rather than accumulating, which is the mechanism behind the context-refresh transition alongside PAGE, REFINE, CLARIFY and STOP.

What is the licensing and backend position?

The current research snapshot intentionally claims no final licensing policy, and the README says to add the repository license before publishing a stable release. On backends it is equally explicit: the runtime API is intended to stay backend-independent, with Jev the first backend used for live experiments rather than the only possible one.

CommunityJev Agents & MCP

Jev Agent Design With Topk Logits Choices

Q: How do I run it without a key or a GPU?

Create a venv, activate it, run pip install -e ., then python -m jev_agent.cli --workspace .\agent_workspace. That core profile needs no model weights and no API key and is described as a deterministic protocol and workspace smoke test, not a language-model quality benchmark. A small optional 0.8B helper for proposing natural-language candidates comes from pip install -e ".[models]", with weights cached outside the repository.

Research runtime that pages a large logical tool and context space through a small bounded Jev decision surface with explicit approvals.

See highlights

Open github.com

Stars: 21
Language: Python
Published by: Community
Author: 6Mikao9
Last verified: Oct 2026

6Mikao9/jev-agent-design-with-topk-logits-choices

User repository on GitHub

View on GitHub

Research design for a Jev-native agent system:enable more options than jev provided with virtulization and paging, tool integration, external helper logits Top-k proposals with Jev-controlled fallback ,decision-aware hierarchical memory, and dependency-aware replanning.

Forks: 2
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

An open research runtime for bounded, non-generative decision models. The runtime lets a model with a finite decision interface operate over a much larger logical space of tools, memories, context blocks, and refinements. Jev is the first backend we use for live experiments; the runtime API is intended to remain backend independent. The core profile runs locally without a GPU, model weights, or an API key. It is a deterministic protocol and workspace smoke test, not a language-model quality benchmark.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Virtual Option Space pages and refines logical candidates through a bounded resident surface.
DecisionModel plus step loop with PAGE, REFINE, CLARIFY, STOP and context-refresh transitions.
File and command tools with approval gates, workspace guards and receipt-based validation.
Dated evidence reports for receipts, verdicts, recovery and live answer-blind paging.

Watch out

Core CLI is explicit rather than autonomous and live runs add remote latency; licensing is still unset. 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-recruiter

Local LinkedIn recruiting workspace on Browser Use's Jev Ultrafast: Jev screens titles, drives navigation, and assesses profile excerpts against your brief while a live browser stays visible; matches and evidence export as JSON.