Jev vs agent loops
An agent loop plans, calls tools, and generates the next message. Jev (TypeSafe System One) does not. Official “how to build”: System One is for AI-powered software, not agents that choose their own next action.
This unofficial comparison is for people arriving from Claude Code / browser-use integration pages. We cover the intent (Jev inside an agent) without cloning those sites’ IA. Not affiliated with TypeSafe or LangChain. No keys sold.
Router, not actor
| Job | Who |
|---|---|
| Type the next user-visible sentence | LLM / template |
| Click the browser / run bash | Agent runtime (Claude Code, computer-use, Playwright) |
| “Which tool is allowed?” / “Which model?” / “Is this injection?” | Jev + your if |
| Control flow | Your code (or the agent host) |
If you ask Jev to be the browser agent, you are off the documented contract (no generation, no images as state unless you transcribe).
Agent-skill pattern (official)
TypeSafe publishes a drop-in skill so coding agents batch questions instead of one HTTP call per Noul:
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai
# or: npx skills add typesafe-ai/skills --skill typesafe-aiRead SKILL.md on github.com/typesafe-ai/skills. The skill teaches the API; it does not replace the agent. Details: agent skill.
Platform-team habits (constants file, review questions, no invented fields): Jev for platform teams.
LangChain depth (net-new vs handbook clones)
LangChain’s 17 Sep 2026 post describes TypeSafeClassifier, ModelRouterMiddleware, and AutoModeMiddleware (tool gates). Those are harness hooks. Jev still only returns typed answers. Vendor speed/cost quotes in that post remain vendor-reported. Walkthrough: LangChain harness.
When the agent loop still wins
- The task is generation + tools (coding, browsing, multi-step research)
- You need a conversation, not a snap decision
- Verifiable System Two work (math with a checker) — launch post’s own split
Use Jev inside that loop for cheap structured gates.
Gate the loop, do not become the actor
Sketch: agent host plans and calls tools → before a privileged tool, your code POSTs filtered state + atomic questions to pinned jev-1.13.0 → if on Choice/Noul → tool runs or is denied. Transcribe pixels before they become state. The skill is a drop-in so the agent batches questions; it is not Jev and it does not issue keys.
TypeSafe’s models page lists jev-1.13 at $0.042 per million input tokens (output free) — a vendor claim, not a loop-cost benchmark. Jaggedness still applies inside the loop: do not ask Jev to do the math the tool should do. Schema-safe “allow tool” is not “the bash command is safe.” Jev does not generate the next user-visible sentence.
How an agent-loop integration eats the contract
- Jev as the browser. No image input. No clicks. Transcribe, then ask, then let Playwright act.
- One HTTP per Noul. Agents default to chat-shaped tool use. Batch questions. Cite the install; do not treat
SKILL.mdas something jev.pro hosts. - Control flow in criteria. The host owns the loop. Do not encode “then search then write” as a Choice.
- Privileged act on timeout. Fail closed. A 529 is not permission.
When not to use Jev here
Do not replace the actor with Jev. If the work is conversation, browsing, or checked System Two math, keep the loop and put Jev only on the gate. The bullets above this section are the cases where the loop still wins.
Criteria that decide the architecture
| Axis | Jev (System One) | Agent loops |
|---|---|---|
| Next user-visible sentence | No generation | LLM / template |
| Tool / browser / bash | Your if after a gate |
The runtime |
| Typed allow / route / inject? | Choice / Noul / Score | Prompt-only is weaker |
| Control flow | Never | Host / orchestrator |
What this page does not claim
- Not a Claude Code or LangChain review.
- Schema-safe output is not the same as factual correctness.
- No independent bake-off numbers live here.
FAQ
Browser-use + Jev?
Transcribe the page (or DOM text) into state, ask Nouls/Choices, let the computer-use agent act. Jev never “sees” pixels.
Is Jev just a smaller LLM? TypeSafe argues no (RLCD, parallel sampler, no strings). We do not run that bake-off here.
Does installing the skill sell me a key?
No. The runtime still needs a documented TYPESAFE_API_KEY or gateway secret. jev.pro does not issue keys.
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.
Hub: Comparisons. Siblings: LangChain harness, vs single-model serving, agent skill. Canonical: https://docs.typesafe.ai.