Does Jev replace the LLM here, or just score it?

It judges while the LLM writes. Gemini 3.8 Flash generates candidate actions during expansion; Jev's Choice returns normalized probabilities used as the policy prior and picks a branching factor, Score gives each hypothetical state a value on a 1-10 rubric, and Noul answers whether an executed command matches the action and whether the goal is complete. Jev never generates the actions.

Will it actually run commands in my workspace?

Yes, but only the single best immediate action, in the workspace you designate, with returncode, stdout and stderr fed back in, and git status, new files and diffs folded into the next round's state. Interactive mode lets you approve, skip or redirect a state before it executes. Workspace execution is optional and mock mode is enabled by default.

Why would you branch by prime numbers?

Choice samples the expansion width from 2, 3, 5, 7, 11 and 13 according to state uncertainty and goal complexity, so a narrow goal gets a small tree and an ambiguous one gets a wider one. That is the project's stated design rather than a tuned default, so it is also the knob most likely to need adjusting for your own goals.

Mock mode is on by default — can a first run look green with no provider key at all?

Yes, and that is the trap. demo --mock runs the whole closed-loop search loop on mock primitives with no network calls and no keys, and the web UI enables mock mode by default too, so a first run can look successful while nothing was really judged. For live work remove --mock, set TYPESAFE_API_KEY, and supply either the Antigravity CLI or a Gemini API key; the loop then ends when a Jev confidence reading reaches 0.85, which is a threshold on a judge the project itself calls heuristic.

What does each Jev primitive do in the MCTS agent?

Noul handles execution checks and early stopping, Choice assigns action priors and picks dynamic branching widths from primes, and Score evaluates states on a continuous 1-10 rubric as a value heuristic.

How does the MCTS agent decide the goal is done?

A Noul question asks whether the goal is complete: yes at confidence 0.85 or above finishes the run, while below that it runs the next MCTS step.

What is the agent's closed-loop execution cycle?

Plan and Choose, Execute, Review, Adapt, Assess; only the immediate winning action is executed in the workspace, and git diffs and terminal observations are folded back into context.

CommunityJev Practices & Patterns

Mcts Agent

Python agent that runs Discriminative MCTS with Jev judging and Gemini 3.8 Flash proposing. Choice supplies action priors and branching width, Score evaluates states on a 1-10 rubric, and Noul checks execution and ends the loop at 0.85 confidence, with a local web UI for following runs.

See highlights

Open github.com

Stars: 7
Language: Python
License: MIT
Published by: Community
Author: lhemerly
Last verified: Oct 2026

lhemerly/mcts-agent

User repository on GitHub

View on GitHub

Discriminative Monte Carlo Tree Search using System One and Harnesses

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

An autonomous reasoning and execution agent powered by Discriminative Monte Carlo Tree Search (MCTS). The system couples TypeSafe Jev System One Primitives (Noul, Choice, Score) for lightning-fast, structured discriminative evaluations with Gemini 3.8 Flash for action generation in a closed-loop execution environment. Includes a local MCTS Agent Web UI for prompting the agent, following its output and image artifacts live, and loading saved run logs.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Score evaluates states on a continuous 1-10 rubric, replacing costly random Monte Carlo rollouts with a discriminative value estimate.
Choice supplies the policy priors and samples expansion widths from 2, 3, 5, 7, 11 and 13 based on uncertainty and goal complexity.
Traversal uses PUCT with first-play urgency, so unvisited children inherit the parent's value instead of starving at zero.
Noul gates execution checks and early stopping, and the loop finishes when the goal-complete confidence reaches 0.85.

Watch out

The README digest available for this review is truncated partway through the usage section, so research mode and anything after the CLI examples were not checked; some badges in the README also point at a different GitHub org than this repository. The System One judgments are explicitly heuristic by the project's own statement, and a Jev confidence reading above 0.85 is what ends the loop, which is a threshold on a heuristic judge. Low early scores deliberately do not prune proposed paths, so trees can stay wide, and mock mode is on by default, which means a first run can look successful with no provider key at all. 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-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.