"10 Levels of Jev For Agentic Engineers" — IndyDevDan

Full transcript: [[2026-09-28-indy-dev-dan-10-levels-of-jev-transcript]].

Why this is in the vault

IndyDevDan is a tracked tier-1 channel for RDCO's agent-harness practice, and Jev/TypeSafe is already an actively-tracked topic in the vault (10+ prior entries in the last two weeks, plus an installed typesafe:typesafe-ai skill and a completed security review of the TypeSafe skills plugin). This video is the first entry to lay out a concrete, ordered decision framework — "when does a narrow classifier replace an agent call, and when does it just support one" — that's directly reusable for RDCO's own gate/guardrail design.

Episode summary

Dan walks through 10 escalating use-levels of Jev (TypeSafe's "System One" fast/cheap JSON-programmed decision model), starting from a simple yes/no classifier (prompt-injection detection) through composite scoring, confidence-gated bash-command blocking, model/agent routing, in-agent guardrail hooks, self-compaction triggers, and finally "agentic Jev" — where the coding agent itself decides when and what to ask Jev. Screen-recorded live demos run each level against a real Pi coding agent, with cost comparisons showing Jev at a fraction of a cent versus dollars-to-thousands for calling a frontier LLM at the same volume.

Key arguments / segments

[00:00:00] Frames Jev as an addition to agent tooling, not a replacement for LLMs — "it's Jev AND your agents," each doing the work they're suited for.
[00:00:57] Level 1 — basic yes/no decision-making ("a smart, cheap, fast if statement"), demoed on prompt-injection / query-injection detection with confidence intervals per input, sub-1-second response time.
[00:03:18] Level 2 — multiple-choice classification, demoed on support-ticket triage (bug vs. not, priority level) with cost comparisons: Jev vs. Gemini Flash (44x), GPT-class models (80-300x), and top-tier models (600x) more expensive at scale.
[00:06:20] Level 3 — composite scoring across founder-defined weighted criteria (e.g. code-review risk: security risk, blast radius, template adherence), tunable by adjusting weights, not the query.
[00:08:50] Level 4 — confidence gating as a credibility control: bash-command gateway demo classifies command reversibility (git push --force = 0.99 irreversible) to decide whether a pre-tool-call gate should block execution.
[00:12:11] Level 5 — routing: Jev picks the cheapest sufficient model or the right specialized agent (browser agent vs. fast agent vs. desktop agent) for an incoming task, still "totally cheap" even at scale.
[00:14:31] Level 6 — in-agent guardrail hooks: live demo of a "Jev guard" pre-tool-call hook blocking rm -rf node_modules, git push --force origin main, and protected-file writes inside a real Pi coding agent, regardless of which frontier model is driving it.
[00:17:34] Level 7 — agent auto-compact trigger: reuses last week's self-compacting Pi agent pattern, but hands the "should I compact now?" decision to Jev (soft/recommend/forced thresholds) instead of a fixed context-percentage default.
[00:20:02] Level 8 — "dirt cheap file reads": Jev answers yes/no/classification questions about a file's contents (does it validate tokens? contain real credentials?) without the file ever entering the coding agent's context window — Dan calls this a genuine mindset shift in how he builds with agents.
[00:23:39] Level 9 — same file-question pattern scaled across many files / a recursive glob in parallel, letting the agent only read the files Jev flags as relevant — demoed finding two relevant files across an entire repo for a rounding-error bug.
[00:29:28] Level 10 — "agentic Jev": the coding agent itself chooses when and what to ask Jev (classify a test failure, verify a fix, assess risk) rather than the engineer hard-coding the call sites.

Notable claims

Jev pricing claimed at roughly 4x DeepSeek V4 Flash and 44x-600x cheaper than mainstream frontier models (Gemini Flash through top-tier "Fable 5.1"-class models) at volume — e.g. $20 for a million calls vs. a claimed $11,000 for the same volume against a frontier model. [00:04:00]-[00:14:00] (Dan's own cost framing from the TypeSafe launch post; not independently verified against published TypeSafe pricing in this note.)
Named pattern: "Jev guard" — a pre-tool-call hook inside a Pi coding agent that classifies bash-command/write-tool reversibility and blocks irreversible or destructive calls before execution, independent of which underlying model is driving the agent. [00:14:31]-[00:17:00]
Named pattern: "cheap file reads" / "files at scale" — routing yes/no or classification questions about file contents through Jev instead of loading the file into the coding agent's context, including recursive-glob fan-out across a whole repo. [00:20:02]-[00:28:00]
Explicit non-sponsorship disclosure: "I am not sponsored. I do not accept any sponsorships on this channel." [00:34:00]
Codebase referenced and shared: github.com/disler/ten-levels-of-jev.

Guests

Solo — no guests.

Mapping against Ray Data Co

Levels 4 and 6 land squarely on RDCO's own harness-engineering practice: the bash-command reversibility gate and the pre-tool-call "Jev guard" hook are the same shape as AGENTS.md's own auto-mode classifier hard-gate (feedback_automode_classifier_hard_gate.md — deploy/production-write denial as a hard stop) and the allow-listed-command discipline already in place for wrangler.sh and other gated scripts. Level 8/9's "cheap file reads without loading the file into context" is a cheaper, narrower cousin of AGENTS.md hard rule #4 (route long artifacts through subagents) — where RDCO currently pays a subagent-dispatch cost to keep context lean, Jev-style classification could answer a yes/no or triage question about a file for a fraction of a cent without any agent dispatch at all, which is worth prototyping against the newsletter/YouTube ingestion pipelines' pre-filter steps. Jev itself is not new to RDCO — the typesafe:typesafe-ai skill is already installed and the topic has 10+ prior vault entries in the last two weeks — so this video's specific contribution is the ordered "level" framework for deciding when a narrow classifier should replace an agent call versus merely gate one, which is a reusable design checklist rather than a new tracked concept.

[[2026-09-23-every-jev-usage-guide]] — most directly related prior Jev coverage; a practical usage guide this video's "10 levels" taxonomy extends into a graduated decision framework.
[[2026-09-21-indy-dev-dan-self-compact-pi-agent-zero-hype-devlog]] — the self-compacting Pi agent this video explicitly builds on for Level 7 (agent auto-compact), referenced directly in the description.
[[2026-09-07-indy-dev-dan-agent-swarms-gpt6-astra]] — referenced directly in the description ("Agent Swarms"); same channel's multi-agent orchestration thread this video's routing levels (5-6) connect to.
[[2026-05-11-indy-dev-dan-delete-bash-tool-agentic-security]] — prior same-channel coverage of bash-tool danger and gating, the direct precedent for this video's Level 4/6 bash-command-gateway demos.
[[2026-08-24-indy-dev-dan-intelligence-explosion-harness-engineering]] — same channel's "harness engineering" framing this video's guardrail-hook and routing levels sit inside.

indy dev dan 10 levels of jev