When an AI agent takes an action it shouldn't, you usually find out too late. Reshimu sits between the agent and the action, and blocks the dangerous ones before they run.
Runtime governance · Irreversible action classification · Intent persistence · Ambiguity escalation
Free during beta · LangGraph · Claude Code · AutoGen · CrewAI · Zero native deps
It already ran. There is no undo. The audit trail tells you what happened, not what to do next.
No single agent is wrong. The session has quietly forgotten what it was for.
One of them will pick the wrong one. Probably the irreversible one. Probably this week.
SHOR extracts every addressable entity from the agent's output and checks each one against the context the agent was given. Below: a partial hallucination caught before it reaches a tool call.
import { classify } from '@reshimu/shor' const result = classify({ output: 'Q3 revenue was $4.2M from 47 customers.', context: 'Q3 numbers: 47 customers signed up, but revenue was not disclosed.', })
{
level: "PARTIAL",
score: 0.6666666666666666,
flagForReview: true,
explanation: "number '$4.2M' not found in context.",
entities: [
{ text: "Q3", type: "date", found: true },
{ text: "$4.2M", type: "number", found: false },
{ text: "47 customers", type: "number", found: true },
]
}
A vertical hierarchy. Intent flows downward through five stages, each one binding the work more tightly to the original instruction. Guardians sit at every level.
Define what the session is for in three bullets. Every downstream agent is scored against it. Drift is caught early — before twenty messages compound into the wrong outcome.
Four guardian classifiers run on every message before delivery. Sub-5ms. No LLM calls for classification. Deterministic where it has to be.
Every intercept. Every decision. Every escalation. Written to disk. Queryable. Exportable. The impression that keeps the system honest.
Detects when agents exceed their assigned domain. Catches scope creep before it propagates downstream and contaminates sibling sessions.
npm install @reshimu/aryehDetects ungrounded confidence and hallucinated claims. Every output must trace to its source. Confidence without basis is a flag, not a result.
npm install @reshimu/shorClassifies all proposed actions GREEN / YELLOW / RED before execution. Blocks dangerous calls automatically. Human-in-loop for red. Blocked
npm install @reshimu/nesherDetects genuine gray zones. Files a structured escalation report. Halts the work until a superior agent — or a human — resolves it.
Panim Adam — פָּנִים אָדָםAll four available on npm.
Be early.
Developer beta · Free to start · No credit card
Reshimu.ai also builds custom AI automation for businesses ready to move beyond experiments. We design, deploy, and govern the systems — so you get the leverage of AI without the risk of agents acting without intention.
Identify the right places for AI. Two-week engagement. Deliverable: full AI roadmap and risk-ranked automation map.
Designed and deployed by us. Always governed by Reshimu's runtime layer from day one. 4–12 week engagements. Production systems, not demos.
We become your AI infrastructure team. Ongoing. For companies treating AI as core infrastructure, not a feature flag.