Hermes Plugin Hands Agent Triage to Jev in 0.4 Seconds
A new open-source toolkit wires TypeSafe's tiny decision model into Hermes agents, handing off routing, memory filtering and skill picks in under a second.
- Kerpopule open-sourced hermes-jev-skills, wiring TypeSafe's Jev decision model into Hermes, Claude Code and Codex.
- Six skills cover model routing, memory filtering, compaction, skill selection, computer use and browser use.
- Each decision returns in ~0.4s with calibrated confidence, for a tiny fraction of a cent per call.
- MIT license, stdlib-only Python 3.9+, no dependencies; installer auto-detects supported agents.
- Everything fails open: outages or low confidence keep the current model and never block a turn.
- API keys stored in OS keychain via one-time local page; only redacted decision inputs leave the machine.
Hermes plug-in delegates agent triage to Jev
An MIT-licensed repository from GitHub developer kerpopule integrates TypeSafe’s hosted Jev decision model with Nous Research’s Hermes coding agent. Named hermes-jev-skills, the project delegates recurring classification tasks such as model routing, memory filtering, context compaction, skill selection, and interface actions. Its plain SKILL.md packages also support Claude Code and Codex.
Agent loops make many small decisions during each turn, and sending every choice to a frontier model adds latency and inference cost. Jev accepts state plus typed questions such as “choose one,” “score this item,” or “answer yes or no,” then returns a decision with a calibrated confidence score. The project reports responses around 0.4 seconds for common routing tasks at a fraction of a cent.
Six jobs for a small decision model
| Agent task | Jev’s role | Reported performance |
|---|---|---|
| Model routing | Select the lowest-cost model suitable for the current turn. | About 0.4 seconds per turn |
| Memory filtering | Score retrieved passages for relevance and flag possible prompt injection. | Up to 60 passages per request |
| Compaction | Choose which prior turns remain verbatim, receive summaries, or leave the active context. | 71 turns in about 0.95 seconds |
| Skill selection | Choose an applicable installed skill from the available catalog. | 373 skills in about 0.9 seconds |
| Computer use | Select the next GUI action from an approved action table. | About 0.4 seconds per step |
| Browser use | Select the next page action from an approved action table. | About 0.4 seconds per step |
The repository supplies these measurements as implementation examples rather than independent benchmarks. Actual latency will depend on network conditions, service load, payload size, and the surrounding agent runtime.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
