ROUTEAgents & automation
Judgment Router for AI Agents
upasana pati@upasanapati
This project implements a System-1 judgment router using `jev-1.13.0` to handle routing decisions before invoking a System-2 LLM, significantly reducing latency and LLM calls.
TakeA `JudgmentSwitch` node (`jev-1.13.0`) evaluates the prompt in one 260ms pass using calibrated `Choice`, `Noul`, and `Score` probabilities before `gemini-3.8-flash` ever wakes up.
If your AI agent calls an LLM on every turn just to pick a route or validate a prompt, you are paying a 10x latency tax!
Here is how we added a 260ms System-1 judgment router (`jev-1.13.0`) in front of Google ADK and cut our Gemini calls by 95 percent. (bookmark it)
Standard single-LLM agents burn 2 to 10 sequential model calls per turn on live MCP tools. They take 40+ seconds on multi-ticket triage, miss hidden SLA breaches, and fire blind searches on vague prompts.
Most of what an agent loses in speed and reliability, it loses by forcing a System-2 LLM to make routing decisions.
The fix routes control flow out of the generative LLM.
A `JudgmentSwitch` node (`jev-1.13.0`) evaluates the prompt in one 260ms pass using calibrated `Choice`, `Noul`, and `Score` probabilities before `gemini-3.8-flash` ever wakes up.
Vague prompts and junk writes abstain at the edge, single-ticket lookups go straight to Cloud Run MCP, and batch triage scores 24 criteria in one `JudgmentMap` call.
LLM calls drop by 95.1 percent, SLA triage runs 3.4x faster (40s to 10s) with 100 percent recall, and gated turns run 15x faster at zero LLM token cost.
Repo: github.com/upasana1105/it-helpdesk-assista
Framework: github.com/mbonnardot/judgment-base-agent
Related builds
