omunculus: teams of AI agents with rules and oversight
Open-source engine that coordinates several AI agents: it hands out work, limits what each one may do, calls a person when needed, and records every step.
- Service
- AI agents
- Area
- IT
- Repository
- github.com/This-Is-NPC/omunculus ↗
01The problem
Multi-step AI work had no single record of who did what. Permissions were loose, and a stuck agent simply stopped, with no retry and no escalation to a person.
How the solution works
-
01
Receive workWorkflow builder · sets the rules
-
02
Split into tasks
-
03
Run isolatedAI agents · run the tasks
-
04
Ask approvalPerson in charge · approves or takes over
-
05
Retry
-
06
Record
02What we did
We built an engine that hands out work items and splits them into subtasks for other agents, runs several agents at once in isolated environments, and applies layered permission limits: each tool is allowed, allowed on request, or blocked. It has a human approval inbox, automatic retries with escalation to a person, and a full event log.
03What it changes
Whoever builds AI agent workflows controls what each agent may do, has a full history to audit, and gets a person stepping in when needed.
More cases
An AI assistant that runs inside the company
Employees talk to the company's documents without anything leaving its servers, and each person only sees what they are allowed to see.
View case →omashiki: work delegated to AI agents, with control and review
Open-source platform that takes tasks, runs each one with AI agents in an isolated environment, and returns a result ready for review.
View case →Mori: factory cameras that warn you when something is off
Uses the security cameras already in place to detect risky situations and alert the team within seconds.
View case →