Vibeleaderboard
Index / tool
Visit github.com
Category
AI Agents
Rank
No. 3761Tools index

Previous survey · No. 3560 ·

Pricing
Open Source
Type
TOOL
Use case
Security & Identity
GitHub
33 stars
Latest release
v0.6.4
Date

About

Submilli is a code-execution runtime with a semantic permission model for business agents that generate code. It enforces rules such as refund limits that differ by customer tier, which prompt-based safeguards only follow some of the time. Running agent-generated code in the runtime lets the agent owner set those limits in the execution layer.

At a glance

Submilli runs agent-generated code under policies that can restrict business actions by their arguments. It is worth evaluating if an agent needs to make calls within limits such as a maximum refund for each customer tier.

Good fit

Teams that can define permitted operations in advance and route them through the runtime.

Consider something else when

Use a simpler fixed workflow when the agent does not need to generate code or choose among actions.

Start here

Define a fake refund tool and test amounts just below and above a limit. Verify the blocked call never reaches the test backend.

Cost and requirements

The Apache-licensed project documents a runtime and policies called Blueprints. You need to define the allowed tools and their rules.

Limits to know

A runtime rule can only cover operations routed through it and conditions encoded in the policy. Review direct access to the underlying services separately.

Tags

agent-runtimepermissionssandboxcode-executionguardrailsrust

Tech Stack

RustDockerTypeScript

Comments (0)

No comments yet

Editorially curated, with community endorsements as a secondary signal. Corrections welcome.