Swarm

data-ai MCP Server

SWARM: System-Wide Assessment of Risk in Multi-agent systems

Verified
data-aidata-ai
3 views29 stars4 forksMIT

Why This Matters

Discovered via agent-topic:ai-agent and last synced 4mo ago.

Verified
Source
agent-topic:ai-agent
Stars
29
Last synced
4mo ago
Install
Check source

Install

Install instructions not detected yet

Check the source repository for the latest setup steps.

View source instructions
14
Tools
0
Resources
0
Prompts
Standard I/O
Transport

Available Tools (14)

Limited

Limited

No

No

Topic

Description

License

MIT

MIT

Varies

Type

Behavior

Feature

SWARM

AgentBench

METR

Custom

Benchmark suites

Example

Description

None

None

Intermediate

> **Tip for Colab users:** All notebooks with a Colab badge auto-detect Colab and install SWARM from GitHub. For scripts without a notebook, add `!pip install swarm-safety` in the first cell. ## Core Concepts ### Soft Probabilistic Labels Instead of binary labels (good/bad), interactions carry a probability `p = P(v = +1)` representing the likelihood of a beneficial outcome: 1. **Proxy signals** are combined into a raw score `v_hat in [-1, +1]` 2. **Calibrated sigmoid** converts to probability: `p = 1 / (1 + exp(-k * v_hat))` ### Four Key Metrics

Formula

Interpretation

Yes

Yes