Artificial Intelligence

Agents can decide without generating text

Weitekamp argues many agent tasks need ranked choices, not generation, using Doom, Factorio and 12,000 categories.

StartupHub.ai Staff

By StartupHub.ai Staff, reviewed by Daniel Singer, Editor

2 min read

12,000 Shopify product categories can be searched without a beam search, Raymond Weitekamp of OpenProse told AI Engineer.

Agents can decide without generating text - AI Engineer
Agents can decide without generating text, AI Engineer

Weitekamp said the trick is recursive decision models. A parent decision calls a child decision up to a depth limit, with context held outside the model in code. He described TypeSafe Jev as the System One counterpart to generative models, because Jev cannot generate text and only ranks options while code makes the decision.

Two coupled Jev compositions played Doom, Weitekamp said. One handled moment to moment shooting choices such as hold, carry, walk or dodge. The other handled map aware exploration, and together they completed levels with no language model in the loop.

He showed the same pattern in Factorio. The system scanned production lines, isolated the failing inserter group and dispatched a fix, all as many small ranked choices. His three rules were that context stays outside, a decision can call another decision, and the tree must first work with zero generative calls before it is allowed to compose with models or agents.

Tool use arrived fast. People built Jev routers, compaction judges and tool gate checks within 24 hours of access, Weitekamp said. The constraint is enumeration, since someone must list the options in code, though he argued a browser page often exposes only three actionable elements in his demo and therefore stays tractable.

Outside that talk, the field is already sorting into decision models versus generators. A community Awesome-Jev list tracks projects built on what it calls TypeSafe AI's System One model for typed decisions, per that listing, while AIMultiple contrasts Jev with Clef and full LLMs and says for most browser steps the decision model reads the page and options, scores each and returns the most likely one. Weitekamp gave no latency, cost or accuracy figures for recursive trees versus single reasoning calls, which leaves the comparison unpriced. Published error and cost curves would change the read. How many real tasks stay solvable when every option must be listed first?

Mentioned in this story

Markets

ShopifyNYSE:SHOP

Sources

Get the free daily brief every morning

Every morning the day's AI news, new AI startups and new funding rounds. Free and takes a few minutes to read, unsubscribe at any time.

Join 64,000 other subscribers.

TopicsRaymond WeitekampOpenProseTypeSafeJevShopifyDoomFactorioClef

StartupHub.ai Staff

Written by

StartupHub.ai StaffEditorial team

The staff writers of StartupHub.ai, ranging from investment analysts to avid AI tool users, early adopters and critical enthusiasts. Backgrounds span engineering, business and the arts. We hold every piece to rigorous standards of research and review.

Daniel Singer

Reviewed by

Daniel Singer is an AI expert who has been writing about, covering and building AI technology since 2018.

Test your Agent Readiness now

Paste your website. How well ChatGPT, Claude and the other agents can read and use a site, in about a minute.

Watch

The videos behind the news, one flip at a time

The feed mixes the videos from our articles with startup demos and you flip through them like TV channels for free with no signup.