Workflow guide

The Dual-Speed Agent: Coupling Codex with Jev for Fast Harness Decisions

Published: 2026-09-23

Decouple deep reasoning from instant decisions. Use Codex as the engineer for architecture and code, and Jev as the fast evaluator, router, and risk gatekeeper.

Dual-speed agent architecture comparing heavy monolithic agent with Codex plus Jev decision layer
  1. Step 1

    Recognize the monolithic agent bottleneck

    Coding agents often send every trivial yes/no question, permission check, and test verification to their heaviest reasoning model. This wastes token budget and adds seconds of latency to loops that should be instantaneous. In a dual-speed system, heavy LLMs like Codex handle deep reasoning, while Jev handles instant micro-decisions.

  2. Step 2

    Route tools with Choice instead of asking for prose

    When an agent needs to pick the best tool among a known set (e.g. ripgrep vs browser vs terminal), do not ask for a paragraph explaining why. A single Choice question returns the selected tool identifier in milliseconds without token overhead.

    {
      "model": "jev-latest",
      "state": "Goal: Locate PlayerHealth definition in the codebase.",
      "questions": {
        "next_tool": {
          "type": "choice",
          "instructions": "Select the best tool for this lookup.",
          "criteria": {
            "grep": "Codebase search (ripgrep)",
            "browser": "Web browser",
            "runtime": "Launch executable",
            "git_log": "Git history"
          }
        }
      }
    }
  3. Step 3

    Intercept destructive bash commands with a Risk Gate

    Before executing terminal commands like git reset --hard, run a Jev Choice or Score risk gate. Low-risk operations execute automatically, medium-risk actions prompt the user, and dangerous commands are blocked before touching disk.

    {
      "model": "jev-latest",
      "state": "Pending command: git reset --hard origin/main",
      "questions": {
        "risk_tier": {
          "type": "choice",
          "instructions": "Classify operational risk of this command.",
          "criteria": {
            "routine": "Safe or read-only",
            "review": "Requires user confirmation",
            "dangerous": "Destructive or irreversible"
          }
        }
      }
    }
  4. Step 4

    Verify test suites with a zero-latency Noul Evaluator

    After running npm test or pytest, pass the exit code and summary to a Noul question. If probability of clean pass is high, proceed immediately to git commit. If it fails, escalate back to Codex for root-cause diagnosis.

    {
      "model": "jev-latest",
      "state": "npm test returned exit code 0. 128 passed. 0 failed.",
      "questions": {
        "tests_passed": {
          "type": "noul",
          "instructions": "Did the test suite finish with zero failures?"
        }
      }
    }
    Closed-loop dual-speed agent engineering workflow
  5. Step 5

    Benchmark: Compare latency and deterministic reliability

    Compared to standard LLMs that generate 450+ tokens of conversational reasoning over 3,500ms, Jev responds in 18ms with 0 token fluff and strict typed outputs. This eliminates Markdown JSON extraction failures in autonomous loops.

    Benchmark comparison showing 99% lower latency and 0 unnecessary tokens
  6. Step 6

    Embed the division of labor in AGENTS.md

    Do not manually prompt the agent on each turn. Document the boundary in your project memory: instruct Codex to use Jev for pass/fail checks, tool routing, and risk gating, while reserving Codex for code authoring, multi-file debugging, and architecture design.

    ## Jev Routing Rules
    Use Jev for decisions that do not require generating prose:
    - Test pass/fail verification
    - Tool selection from a known set
    - Terminal command risk classification
    - Continue vs retry vs halt gates
    
    Keep Codex focused on:
    - Writing and refactoring code
    - Multi-file debugging
    - Architectural planning

Related projects

  • fast-jev-compaction — Claude Code plugin and npm library that asks Jev which tool calls to keep, instead of summarizing the transcript.
  • jev-sentinel — Pi, Claude Code, and Codex CLI guard that asks Jev whether a tool call is on-task, risky, or injected before it runs.
  • @typesafe-ai/sdk — Official TypeScript/JavaScript client for POST https://api.typesafe.ai/v1/systemone.

← All guides · Jev Playground