Releases·Tool Release
Jive: open-source terminal coding agent replaces tool calls with DAGs of tool and System One model calls
A developer released Jive, an open-source terminal coding agent that replaces the standard LLM-call-to-tool-call agentic loop with 'graph calls' — each graph is a DAG combining tool calls and 'System One' model (Jev) calls for fast, intuitive decisions. The author benchmarks Jive against Codex and Claude Code on six tasks, reporting large time and token reductions. The post frames the design as letting an agent natively 'think fast and slow' instead of mimicking it with workflows or multi-agent architectures.
18 engagement1 sourceMon, Sep 21, 2026, 08:53 PM
AI Overview
- Jive is an open-source terminal coding agent whose graphs execute tool calls and Jev System One model calls, aiming to let agents "think fast and slow" natively
- Author's benchmarks show Jive finishing six tasks in 1m41s–3m26s and 5.7k–11.1k tokens versus Codex and Claude Code at medium effort
- Jive suits repo investigation, bulk classification, multi-step profiling, and repetitive edits, but the author says it does not yet replace Codex or Claude Code
Entities
Related
Products·launched
Chatlo – iOS-native AI agent app, now with Jev integration
Chatlo is an iOS app that runs AI agents entirely on an iPhone, without needing a separate computer or server, similar to desktop agents like Claude Code or Codex. Its latest version integrates Jev, described as a decision-making AI, which the maker claims is a first on the App Store. It targets iPhone users who want AI agent capabilities on mobile.
Chatlo is an iOS app that runs AI agents entirely on an iPhone, without needing a separate computer or server, similar to desktop agents like Claude Code or Codex. Its latest version integrates Jev, described as a decision-making AI, which the maker claims is a first on the App Store. It targets iPhone users who want AI agent capabilities on mobile.
1 engagement1 sourcehackernewsThu, Sep 24, 2026, 02:23 AM
Social·Community
I tested Jev against 7 frontier models on a real task. Faster and far cheaper, but none of them predicted what I actually prefer
A developer tested the Jev decision model against Opus 5, Sonnet 5, Haiku 4.5 and four Codex models on scoring 110 YouTube subscriptions for keep/drop decisions. Jev ran in 9.1 seconds at about $0.004, far faster and cheaper than the LLMs, but every model predicted the user's own keep/drop labels poorly (AUC 0.43–0.52) while a simple watch-frequency heuristic scored 0.69. The post questions how to establish trustworthy ground truth on subjective personal tasks.
A developer tested the Jev decision model against Opus 5, Sonnet 5, Haiku 4.5 and four Codex models on scoring 110 YouTube subscriptions for keep/drop decisions. Jev ran in 9.1 seconds at about $0.004, far faster and cheaper than the LLMs, but every model predicted the user's own keep/drop labels poorly (AUC 0.43–0.52) while a simple watch-frequency heuristic scored 0.69. The post questions how to establish trustworthy ground truth on subjective personal tasks.
8 engagement1 sourceredditSun, Sep 20, 2026, 05:29 PM
Social·Community
What’s next after Jev? Metacache: Reasoning by Construction
An article argues the industry is moving from monolithic models toward compound systems like Jev, where small agents execute narrow tasks under a code orchestration layer. It proposes 'reasoning by construction': a model builds and returns a reusable compound system — a 'reasonlet' — alongside its answer. Saved reasonlets ('metacaches') can be rerun or edited locally, or cached server-side to save compute.
An article argues the industry is moving from monolithic models toward compound systems like Jev, where small agents execute narrow tasks under a code orchestration layer. It proposes 'reasoning by construction': a model builds and returns a reusable compound system — a 'reasonlet' — alongside its answer. Saved reasonlets ('metacaches') can be rerun or edited locally, or cached server-side to save compute.
13 engagement2 sourcesredditTue, Sep 29, 2026, 06:47 PM
Developments·Paper
REFLEX with Jev for Efficient Selective Control in LLM Agents
A new arXiv paper presents REFLEX, an LLM agent architecture that uses Jev as a fast, typed decision layer and escalates to a strong LLM only on low confidence or when generation is required. On a frozen 100-task benchmark it reaches 95% success with 72.7% fewer strong-model calls. The paper also delineates limits: gains shrink when simple routing cascades are already accurate.
A new arXiv paper presents REFLEX, an LLM agent architecture that uses Jev as a fast, typed decision layer and escalates to a strong LLM only on low confidence or when generation is required. On a frozen 100-task benchmark it reaches 95% success with 72.7% fewer strong-model calls. The paper also delineates limits: gains shrink when simple routing cascades are already accurate.
0 engagement1 sourcearxivTue, Sep 22, 2026, 02:54 PM
Releases·Tool Release
Mellow Machines Lab open-sources Mellow Harness, an agent harness for Jev-based AI hardware with 227 ms median decision latency
Mellow Machines Lab, a small open-source team, released Mellow Harness, an agent harness for building configurable, interactive AI hardware and companions. A System One Model called Jev sits at the center of the harness loop, producing fast expressive feedback via mood graphs, authored personality, and a shared event log. The team also shipped Buddygotchi, a pixel-animated coding companion demonstrating the harness.
Mellow Machines Lab, a small open-source team, released Mellow Harness, an agent harness for building configurable, interactive AI hardware and companions. A System One Model called Jev sits at the center of the harness loop, producing fast expressive feedback via mood graphs, authored personality, and a shared event log. The team also shipped Buddygotchi, a pixel-animated coding companion demonstrating the harness.
1 engagement1 sourceredditFri, Oct 9, 2026, 03:33 PM