Jev Agent: How Jev Works Inside an AI Agent Loop

curl -X POST https://jev-ai.pro/api/v1/systemone \
  -H "Authorization: Bearer $JEV_AI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-latest",
    "state": {
      "task": "Rotate the staging database password and update the app config",
      "proposed_command": "psql -h db.prod.internal -c \"ALTER USER app PASSWORD ...\"",
      "recent_tool_output": "Found config/database.yml with host db.prod.internal"
    },
    "questions": {
      "touches_prod": { "type": "noul", "instructions": "Does the proposed command act on a production system?" },
      "reversible":   { "type": "noul", "instructions": "Can the effect of the command be undone without data loss?" },
      "matches_task": { "type": "noul", "instructions": "Is the command a faithful step toward the stated task?" },
      "next": { "type": "choice", "instructions": "What should the agent do now?",
                "criteria": { "run": "Run the command", "ask": "Ask the user before running it", "revise": "Rewrite the command first" } }
    }
  }'

The response carries touches_prod.noul, reversible.noul, matches_task.noul and next.choice with a probability per option. We ran this exact request on 2026-10-09: touches_prod 0.92, matches_task 0.12 because the task says staging and the command targets a production host, reversible 0.75, and for next: run 0.00, revise 0.52, ask 0.48. The loop never runs the command; it goes back to the LLM to rewrite it or to the user, which is the branch you want.

// One iteration of an agent loop with Jev at the gate (Node 22+, server-side)
const JEV = 'https://jev-ai.pro/api/v1/systemone'
const headers = { authorization: `Bearer ${process.env.JEV_AI_API_KEY}`, 'content-type': 'application/json' }

async function decide(state, questions) {
  const res = await fetch(JEV, { method: 'POST', headers, body: JSON.stringify({ model: 'jev-latest', state, questions }) })
  if (!res.ok) throw new Error(`Jev ${res.status}`)
  return (await res.json()).answers
}

async function step(task, history, tools) {
  const plan = await llm.propose({ task, history, tools })           // the LLM writes the next action
  const a = await decide(
    { task, proposed: plan, recent: history.slice(-3), tools: tools.map(t => ({ name: t.name, description: t.description })) },
    {
      tool: { type: 'choice', instructions: 'Which tool should the agent call to satisfy the task?', criteria: { ...Object.fromEntries(tools.map(t => [t.name, t.description])), none: 'No listed tool fits' } },
      risky: { type: 'noul', instructions: 'Would running the proposed action delete data, spend money or touch production?' },
      done: { type: 'noul', instructions: 'Is the task already complete according to the history?' },
    },
  )
  if (a.done.noul >= 0.9) return { status: 'done' }
  if (a.risky.noul >= 0.5) return { status: 'ask_user', plan }         // a person confirms anything risky
  if (a.tool.probabilities[a.tool.choice] < 0.7) return { status: 'ask_user', plan }
  if (a.tool.choice === 'none') return { status: 'answer', plan }
  return { status: 'run', tool: a.tool.choice, args: plan.args }       // your code runs the tool
}

Three rules do the work: stop when done is above 0.9, hand anything risky to a person, and only run a tool the model is at least 0.7 sure about. On our test, acting only above 0.9 lifted accuracy from 93.0% to 95.6% and escalated 34 of 440 decisions.

// Official TypeSafe SDK, published by TypeSafe — @typesafe-ai/sdk 0.6.0
// Server-side JavaScript / TypeScript using the Jev AI compatible endpoint
import { TypeSafeClient, choice } from '@typesafe-ai/sdk'

const apiKey = process.env.JEV_AI_API_KEY
if (!apiKey) throw new Error('Set JEV_AI_API_KEY on your server')

const client = new TypeSafeClient({
  apiKey,
  baseURL: 'https://jev-ai.pro/api', // Required, including for existing SDK users
  retry: { maxRetries: 0 }, // Avoid replaying a request with an uncertain outcome
})

// The SDK appends /v1/systemone to baseURL.
const response = await client.systemOne({
  model: 'jev-latest',
  state: { document: 'I was charged twice. Please fix this ASAP.' },
  questions: {
    category: choice('What is this ticket about?', { billing: null, technical: null, other: null }),
  },
})
console.log(response.answers.category.choice)

Setting baseURL is required; changing only the key sends the request to TypeSafe with a key it does not recognise. The documentation covers retries, errors and limits; where to run Jev lists OpenRouter, Vercel AI Gateway, Cloudflare, Bedrock, Vertex and Azure for loops that already live there.