WHAT NEW IN X11.06 FRI

AI collections from X · 23 collections · 69,408 posts

jevVIDEO

10 open-source Jev combos grouped by Choice, Score and Boolean decisions

Pairs Jev with MCP SDK, LangGraph, Qdrant, Temporal, Langfuse, Promptfoo, PostHog, Semgrep, Playwright and Stripe, giving reusable patterns by decision type.

Asteri
@Asteri_ethArchived 2026.09.28Originally posted 2026.09.27
Asteri @Asteri_ethOriginally posted 2026.09.27
Open original on X

your agent probably burns a frontier model to pick between 3 buttons 10 open-source Jev combos, grouped by the decision it actually makes. steal the pattern, swap the tool Jev returns Choice, Score or Boolean. one bounded judgment. code owns the permission, the threshold and the fallback WHEN THE ANSWER IS ONE OPTION FROM A KNOWN SET (Choice) 01 Jev × MCP SDK pick which tool to call from the connected servers ▸https://github.com/modelcontextprotocol/typescript-sdk 02 Jev × LangGraph pick the next node in the graph ▸https://github.com/langchain-ai/langgraph 03 Jev × Qdrant pick which retrieved chunk actually enters the context window ▸https://github.com/qdrant/qdrant 04 Jev × Temporal pick which workflow branch to take on retry ▸https://github.com/temporalio/sdk-typescript WHEN THE ANSWER IS A POSITION ON A RUBRIC (Score) 05 Jev × Langfuse score each trace on your own rubric, log it next to the span ▸https://github.com/langfuse/langfuse 06 Jev × Promptfoo score outputs in eval runs without writing a judge prompt from scratch ▸https://github.com/promptfoo/promptfoo 07 Jev × PostHog score inbound events for triage before they hit a human queue ▸https://github.com/PostHog/posthog WHEN THE ANSWER IS TRUE OR FALSE UNDER A POLICY (Boolean) 08 Jev × Semgrep does this diff need a human reviewer before merge, yes or no ▸https://github.com/semgrep/semgrep 09 Jev × Playwright did the page actually reach the state the agent said it did ▸https://github.com/microsoft/playwright 10 Jev × Stripe is this refund request inside policy or does it need approval ▸https://github.com/stripe/stripe-node the loop stays the same for all ten: read the state -> ask one bounded question -> check confidence in code -> run the allowed action or escalate -> verify the real outcome -> log everything architecture patterns, not official integrations the cheapest model call is the one code never had to make

This post quotes another post: x.com/xmyttle/status/2103904144289390899

x.com/asteri_eth/status/2104171216655274431Archived 2026.09.28 10:03