Ray Amjad shows how TypeSafe’s Jev could give Claude Code and Codex a cheap, fast decision layer. The video mixes hands-on demonstrations, other developers’ reported results, and ideas he still plans to test—not a controlled benchmark establishing the title’s “cheapest” claim.
A classifier, not another text generator Jev takes a prompt, data, and predefined answers, then returns probabilities rather than writing a response token by token. Amjad calls it a smart switch statement: ordinary code decides what happens after the model scores the options. Bool evaluates a yes/no question. Choice selects among named options. Score places an input along an ordered, explicitly described rubric. His invoice demo shows why criteria matter: adding specific fraud signals changes the score. Repeated calls produce similar, but not identical, probabilities. His scoring examples cover sales-lead quality and log severity. He describes up to 255 choice options and up to 11 rubric positions for scoring in the version demonstrated. Several questions can share an input: a code diff can be checked for task completion, weakened tests, verification quality, and risk in one request. Fast decisions and deliberate planning do different jobs Amjad uses “System 1” for fast, reflex-like classification and “System 2” for slower planning and reasoning. The point is to combine them, not replace the latter. A fast layer can make frequent bounded decisions; a reasoning model can review outcomes and revise its criteria, examples, or thresholds. His trading-bot example illustrates low-latency action selection, not a demonstrated profitable strategy. He explicitly discourages treating it as a trading recommendation. The integrations discussed use TypeSafe’s service and an API key. Running a controller or browser on a local machine does not mean TypeSafe’s Jev weights are running locally. Minecraft makes the division of labor visible Amjad sets the overall goal: build shelter and obtain a diamond pickaxe. A reasoning model plans the strategy and reviews progress every two minutes and after significant setbacks or milestones. Jev chooses immediate tasks and tactics from the current state: health, hunger, time of day, mining progress, recent events, and available actions. A controller mod performs the physical movements. The run progresses through shelter and tools, pauses to survive the night and heal, obtains a diamond pickaxe, and eventually reaches the Nether. This is a layered agent demonstration, not Jev independently inventing a plan or controlling the game without supporting software. Select skills before filling the context window Large skill libraries can consume substantial context just by exposing every description. Amjad proposes using Jev to identify relevant skills before loading them into the main agent. He cites TypeSafe’s skill-suggestion cookbook: for a Hermes setup with 182 skills and Haiku 4.5, the reported wrong-skill rate falls from 17% to 7.3% with Jev suggestions. His proposed benefit is less irrelevant material in the main model’s context, as well as better routing. Those percentages belong to the cited example, not every agent or skill library. Browser verification as a cheap feedback loop Amjad highlights a browser-use demonstration reported to find a flight in roughly seven seconds for about 0.4 cents. He uses it to motivate rapid checks of application user flows. The proposed loop is: a coding agent implements a feature, a Jev-powered browser agent checks it, and the reasoning model fixes the failures. The reasoning model can also revise the browser agent’s options and criteria. He discusses another team’s parallel adversarial-browser testing and imagines running many such checks per pull request or continuously. Those large-scale testing costs are projections and third-party reports; he says he intends to test the approach himself. Browser and sandbox compute may become the limiting expense even when model calls are cheap. Audit comments before paying an agent to rewrite them Comment accuracy and usefulness are separate: “multiply by two” can accurately describe a line while adding no useful explanation. Amjad supplies the TypeSafe skill and asks his coding agent to propose criteria, obtain his approval, run a small sample, and estimate the full cost before expanding the work. The displayed sample analyzes 150 comments in 9.3 seconds for about one cent. The agent estimates roughly 57 cents to analyze the whole codebase and a shortlist of around 1,700 comments to improve. Classification creates the shortlist; separate coding agents would do the rewriting. The full-codebase figures are estimates, not a completed rewrite demonstrated in the video. Qualitative linters and code-smell screening A qualitative linter asks questions conventional syntax checks struggle to express: does a function name describe its side effects, or does a log statement expose sensitive information? Amjad proposes cheap checks on every pull request, with different handling for secrets, financial values, and personal information. For code smells—design warning signs such as duplication, dead code, or unexplained constants—his agent separates checks suitable for static analysis from questions worth sending to Jev. The agent estimates an exhaustive classification pass at 28 million input tokens and $1.19. Amjad acknowledges that Jev cannot find every kind of code smell. His refinement loop starts small, reviews the classifier’s decisions with a stronger model, revises the rubric, and runs again. This is adjustment of instructions and criteria, not a demonstrated retraining of Jev’s weights. Code review: screen broadly, investigate selectively Amjad points to Jev Review as an example of asking many inexpensive questions about a diff and forwarding important findings to a coding agent. His own agent suggests this could substantially reduce how much a full reviewer must read; the claimed tenfold reduction is a proposal, not a measured end-to-end result in this video. He imagines codebase-specific “reflexes”: invariants, security questions, compatibility checks, and rubrics applied to each pull request. Broad screening would trigger narrower follow-up questions and specialist review. He cites a Sentry engineer’s report of faster, cheaper classification with high accuracy on one security pipeline. That is supporting evidence from a particular workload, not proof that Jev replaces comprehensive security review. The pattern to test Keep reasoning, implementation, and difficult diagnosis with the coding agent. Give the classifier narrow questions, explicit allowed outcomes, and a clear path for escalating findings. Use cheap coverage to identify where expensive attention is worthwhile: skills, comments, suspicious code, risky changes, or failing browser flows. Amjad’s larger prediction is that coding agents will increasingly combine fast decision models with deliberate reasoning models. His concrete starting point is a small measured trial before trusting a broad automated workflow.