Jev in an agentic harness: where it fits in an agent loop

Where Jev sits in an agent harness

A harness is the code around a model that runs the loop: it observes, picks an action, executes it and checks the result. TypeSafe is direct about Jev's role: "System One is TypeSafe's model for building AI-powered software, not agents. It does not generate code or choose its own next action" (How to build with TypeSafe).

In practice, the harness builds the list of possible actions and asks Jev to choose among them. Jev's answer is a probability over the options you supplied, never a new action it made up. The harness then executes, logs and decides what happens next. Anything that needs generated text, such as filling a form field or writing a reply, goes to an LLM. The agents and browsers projects mostly follow this split.

Why Jev fits the decision points

• No invented actions. Every output is limited to the supplied options, so the model returns a full probability distribution "rather than inventing a value outside the schema" (How to build with TypeSafe).
• Knowing when to stop. Choice and Score answers include confidence, so the harness can decide when to act and when to escalate to a person or a reasoning model (System One).
• Parallel checks. All questions in a request run in parallel against the same state, so a step can ask "which action?" and "is this safe?" in one call (Introduction).

TypeSafe's launch post frames the stakes: a hallucinated tool call "is inconvenient in an agent, but is an absolute deal-breaker if it’s part of a system with latency guarantees" (launch post). What is Jev? covers the model itself.

Documented pattern: typed function calling

TypeSafe's function calling cookbook turns a plain-English request into a call to ordinary typed functions. An argument that takes one value from a fixed list becomes a Choice question, a separate yes/no question checks whether the user mentioned that argument at all, and one more Choice picks the function. For ten trading functions, that made 54 questions per command in a single request (Function calling). A sample result from the cookbook:

"plot rolling correlation between nvda and spy for the past month"
    rolling_correlation(symbol='NVDA', benchmark='SPY', window='1mo')   confidence 0.91

The call's confidence is the least certain judgment behind it, not the product of all of them, "since one wrong argument is enough to spoil the result." Arguments that are free text, numbers or dates get no question and keep their defaults (Function calling).

Documented pattern: choosing a skill for each turn

The skill suggestion cookbook targets a common harness problem: agents with large skill rosters see only truncated descriptions and often load the wrong skill. Using the 182 skills of the Hermes agent harness, one Jev request ranks every skill and asks whether the turn needs one at all, and a second re-reads the top three in full and may reject them all. The winner is added as one line to the system prompt (Skill suggestion).

The agent in that test was claude-haiku-4-5-20251001. For more on this job, see routing and model choice.

Gate risky steps on confidence

TypeSafe's confidence-gated routing example sets one floor for every action and a higher bar for risky ones. Below 0.6 confidence the request goes to a person. A balance check runs at 0.6, approving a transfer automatically needs more than 0.85, and a transfer between 0.6 and 0.85 asks the user to confirm (Confidence-gated routing). The same logic works for agent steps: reading a page can run on a lower bar than deleting a file. For screening what goes into and out of the LLM in the loop, TypeSafe also has a guardrails cookbook (Guardrails for LLMs); see safety and guardrails.

Projects doing this

Agents & browsersJev UltrafastBrowser Use's fast web agent, where Jev picks the operation and page element in one request and a small LLM writes text only when typing is needed.GitHub · ★ 20k · browser-useAgents & browsersComputer UseAn experimental macOS computer-use loop that reads the screen with OCR and asks Jev to pick the next click, scroll or text entry.GitHub · ★ 955 · awlevinContext & memoryForemanAn experimental Python supervisor that uses Jev to assess a running Codex worker and deterministic policy to steer, stop, retry, verify or finish the job.GitHub · ★ 554 · thruwireRouting & model choicehermes-jev-skillsAgent skills and a jev CLI that use Jev for model routing, memory filtering, compaction, skill selection and computer or browser actions on Hermes, Claude Code and Codex.GitHub · ★ 726 · kerpopuleAgents & browsersmobile-jevA standalone Android agent for Mobilerun where Jev makes every decision, with a live React studio.GitHub · ★ 380 · droidrunAgents & browsersJevHarnessLLM-authored, task-specific Jev harnesses with optional reward reflection over the full trajectory.GitHub · ★ 191 · TianyuCodingsRouting & model choicejev-dsh-decisionA structured decision plugin for agent harnesses, built on Jev, with native DeepSeek Harness support.GitHub · ★ 163 · Devin-AXISAgents & browsersjev-browser (Ying-Kai-Liao)A JavaScript browser library, CLI and MCP server where the calling agent sets the goal and Jev selects actions from structured page observations.GitHub · ★ 82 · Ying-Kai-LiaoCoding & code reviewjev-useA Claude Code, Codex and pi plugin that hands agent steps needing no text output to Jev and escalates the rest back to the LLM.GitHub · ★ 25 · shitianfangSDKs & integrationssuper-jevA small TypeScript harness that turns evidence into typed Jev judgments, permitted actions and verified outcomes, with local JSONL traces.GitHub · ★ 11 · Kevthetech143Routing & model choicejev-harnessA TypeScript decision harness that maps Jev answers to actions through a policy, a confidence gate and an optional shadow mode.GitHub · ★ 14 · AntonioCoppeRouting & model choicejev-layerA portable decision layer for agent harnesses with host-owned routing over candidate capabilities, receipts and replay, and fail-open integrations.GitHub · ★ 2 · typakon4

Questions

More guides: What is Jev?Jev pricing and API costJev vs LLMs: when to use whichHow to use Jev: a quickstartJev and MCP: using Jev as a Model Context Protocol toolHow to use Jev with Claude CodeJev for orchestration: routing requests, models and skillsJev as a judge: evaluating LLM and agent outputsJev for SEO and GEOJev for ads and ad reviewJev for marketing and lead scoringIs there an open source Jev?Can you fine-tune Jev?Designing questions for JevWhat is RLCD?Jev statistics