JEV_ADK
Python agent development kit for building System One pipelines with Jev for guardrails, routing and PR triage.
Overview
JEV ADK is a Python Agent Development Kit built around the slogan LLMs write, Jev decides, code enforces. It argues that most model calls in agents are micro-decisions, such as whether a bash command is destructive, which tool to call next, or whether a PR should trigger a full regression suite, and that these should go to a fast System One model rather than a large autoregressive LLM. It frames this as a dual-brain design: Jev for fast parallel triage, guardrails and routing, and frontier LLMs for slower reasoning.
The README presents core primitives, a quickstart and production blueprints for a bash guardrail and firewall, dual-brain agent routing, and instant PR triage with a risk matrix. Its illustrative comparison claims around 85 ms and very low cost per Jev decision versus seconds and cents for an LLM, figures that are the project's own. It is MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this proposed bash command destructive? | yes · no |
| Choice | Which tool should be called next? | read_file · bash |
| Score | What is the risk level of this PR? | — |
How it works
- Define state and typed questions with ADK primitives
- Jev answers in one pass
- Code enforces block/allow or routes to an LLM
- Use blueprints for guardrails, routing or PR triage
Strengths
- Ready-made blueprints for common agent decisions
- Clear separation between decision and generation
Caveats
- Benchmark and cost figures are illustrative and self-reported
- Marketing-heavy README; maturity unclear
At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- library
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Agent Development Kit for System-One AI: Sub-100ms non-autoregressive decision…
wh000wh000/awesome-jev-live
Frequently asked questions
What does JEV ADK provide?
Primitives and blueprints for using Jev as the fast decision layer in agents.
What kinds of decisions?
Destructive-command checks, tool selection, agent routing and PR risk triage.
Are the speed figures verified?
They are the project's own illustrative numbers.
Related projects
Pi coding harness where any LLM writes and Jev returns typed reflex scores that policy code turns into allow, confirm, block, warn or steer.
ProjectprototypeTypeScriptLarge catalog of Claude Code agents, commands, hooks and MCPs, listed with community Jev mods for guardrails and subagent models.
CLIproductionJavaScriptCoding-agent hook that asks Jev one yes/no per project rule on every edit or turn and makes the agent fix any violation.
HookbetaTypeScriptGlue that turns one Linux machine into a team of coding agents, with Jev making typed calls like who builds a task and whether it may merge.
ProjectbetaTypeScriptClaude Code plugin sending small judgments to Jev: rule checks that block edits, verbatim compaction, subagent sizing and routing hints.
PluginbetaPythonDesktop workspace for coding agents that runs tasks in parallel worktrees, listed as using Jev to pick the best agent per task.
ProjectbetaTypeScript, Rust