Field guide / primary domain
316
sources in this field
Updated September 20, 2026
Current thesis
The shortest path to orientation.
This synthesis records claims and practices from the cited sources; reported outcomes and product capabilities have not been independently verified.
Production agent stacks deliver ~10x cost reduction on routine tasks through memory, context engineering, skills, and evals. A 300+ bank account reconciliation case study compressed a 12-step workflow (1 month cycle, 12,000+ items) into 3 steps, cutting month-end close from 18–22 days to 7–9 days, AP exceptions from 600–800/month to <50, and deals idle time from 23 to 6 days, with measured client value >$100M. Real waste lives in process design, not task speed. Classify every process step into Deterministic (rule-based code), Agentic (thousands of judgment examples, low risk), or Human-in-the-Loop (agent gathers evidence, human decides in seconds). Frontier advantage shifts to proactive background agents; token-caching strategy demands care: Anthropic cache writes cost 12x reads. Specialized decision models like Jev enable speculative batch querying—adding questions barely changes latency or degrades answers to prior questions, only costs tokens for new ones—and offer 20–200x speed and 40–400x cost improvements for known-answer classification (routing, ticket triage, eval verdicts). A Good Start Labs benchmark across 6,003 rubric checks found Jev matched Claude Fable 5.1 91.5% of the time at $160/million vs $33,000 for Fable 5.1, though open-source DeepSeek V4.1 Flash achieved 93.5% agreement for $260, making the tradeoff less clear-cut. Jev's constraints—no abstention, no reasoning traces, context rot—necessitate single-failure-mode evaluators with explicit true/false criteria. Routing by confidence band (automatic on high confidence, human review on mid, flag low) outperforms forcing a single threshold.
Evidence board
Claims worth carrying forward01Jev evaluates every question in parallel and in isolation against the same state, so adding a 4th or 14th question barely changes latency, only costs tokens for that question, and cannot degrade answers to other questions—enabling speculative batch querying.
02TypeSafe's Jev offers three typed question primitives: Choice (pick 1 of up to 255 options with per-option probabilities), Score (rate against up to 10 ordered rubric levels, returns probability-weighted value + distribution), and Noul (yes/no with probability of true, no separate confidence field).
03TypeSafe claims Jev is 20-200x faster and 40-400x cheaper than frontier models for classification/decision tasks (agent routing, ticket classification, escalation decisions, eval rubric verdicts) where possible answers are known in advance.
04Good Start Labs benchmark: across 6,003 rubric checks, Jev matched Claude Fable 5.1's verdict 91.5% of the time at $160/million graded answers, vs $33,000 for Fable 5.1, $400 for GPT-5.6 Luna, and $1,600 for Gemini 3.8 Flash—suggesting many routing/classification/eval tasks don't need frontier models.
05Open-source comparison caveat: DeepSeek V4.1 Flash cost $260 (vs Jev's $160) but agreed with Claude Fable 5.1 93.5% of the time (vs Jev's 91.5%)—only 2 points better accuracy for $100 more, making the specialized-vs-open-source tradeoff less clear-cut.
06Jev's documented weaknesses for eval design: it cannot abstain (forced binary picks least-wrong answer instead of 'unknown'), gives no rationale for verdicts (debug via your own criteria, not model reasoning), and suffers explicit 'context rot' where accuracy drops as irrelevant state material accumulates—context limits are unclear (docs say 64k/32k, OpenRouter lists 32K).
07Best practice for LLM-as-judge (reinforced by Jev's atomic-question requirement): design one evaluator per failure mode with distinct true/false or category criteria rather than a single vague judgment prompt—forces clearer definition of 'good' and yields calibrated confidence signals.
08Practical pattern: because Jev/Noul-style judges return probabilities rather than binary labels, you can route by confidence band—act automatically on high-confidence verdicts, send mid-confidence cases to human review, and drop/flag low-confidence ones—rather than forcing a single threshold.
Adjacent fields
Latest evidence
Marc Klingen
TypeSafe launched Jev, a non-generative 'system one' model that answers typed Choice/Score/Noul questions with probabilities against a shared state, claimed 20-200x faster and 40-400x cheaper than frontier LLMs for classification/routing/eval-scoring tasks, with documented limits around abstention, rationale, and context rot.
Rafal Wilinski
Rafal Wilinski credits @typesafeai's Jev with building a massively parallel, browser-based adversarial testing suite that stress-tests each release automatically, at very low cost.
Bowe Frankema
Bowe Frankema shares a quick workflow for adopting 'Jev' (a typesafe-ai tool) in existing projects: install the official skill via CLI, then have an AI agent analyze the codebase for integration points.
Cody Schneider
Cody Schneider highlights Graphed.com, a platform letting AI agents run waterfall enrichment across multiple data providers (Findymail, People Data Labs, Prospeo, LeadMagic, Apollo, LeadMarina) on a pay-as-you-go API basis instead of paying for multiple subscriptions. Graphed also offers a broader marketing agent platform with a data warehouse, 750+ source integrations, and prebuilt agents for ads, SEO, content, and outbound.
Heinrich
Heinrich launches the early alpha of Ars Umbris, a malleable IDE that treats a knowledge base like a codebase—typed files, composable repos, cross-repo dependencies, and a type engine that gives agents compiler-like feedback on knowledge structure.
Levi Munneke
A cold email agency architecture uses four persistent GrokBot agents (list builder, copywriter, campaign manager, infra manager) coordinated via a shared thread, with GPT-6 Astra as the coding/data layer, to run cold email GTM as a chain of SOP-defined artifacts with strict handoff rules.
Aaron Levie
Box CEO Aaron Levie argues that enterprise AI value comes from an 'applied layer' that closes the chasm between raw model capability and real workflows—via context aggregation, human-in-the-loop design, evals, and governance—and that this layer becomes more, not less, important as models improve.
Michel Lieben
A GTM agency's detailed playbook for running outbound campaigns through Claude Code as the orchestrator, with Apollo/Clay/Instantly-equivalent tools (via MCP) as endpoints. Covers folder structure, MCP connection setup, a seven-step signal-to-meeting workflow, skills files, reusable prompts, and eight recurring failure modes.
Y Combinator
YC convened researchers and founders to examine how agent 'harnesses' (scaffolding, context management, tooling) drive dramatically different performance from identical model weights, covering harness history, self-improving harnesses, and YC's internal agent stack (QM).
vas
Vas (CEO of Varick Agents) argues enterprise AI spend fails because companies apply AI to broken processes instead of redesigning them, echoing Michael Hammer's 1990 critique of IT investment. He details a mapping methodology (process mining + interviews), a 3-bucket classification for steps (deterministic/agentic/human-in-the-loop), and real deployment results.