autogen-verigent

Trust verification for Microsoft AutoGen multi-agent systems.

The Problem

When multiple AI agents collaborate in an AutoGen GroupChat, you have no way to verify their capabilities or prioritise the right agent for a given task. Agents self-declare what they can do — there's no independent trust signal.

autogen-verigent integrates Verigent trust keys into AutoGen's speaker selection, so agents with verified capabilities get priority for relevant tasks.

Install

pip install autogen-verigent

Quick Start

from autogen import AssistantAgent, UserProxyAgent
from autogen_verigent import VerigentAgent, VerigentGroupChat

# Create your agents as normal
coder = AssistantAgent(name="coder", system_message="You write code.")
reviewer = AssistantAgent(name="reviewer", system_message="You review code.")
user = UserProxyAgent(name="user")

# Wrap with Verigent identity
vg_coder = VerigentAgent(
    coder,
    "VG:CODER-01:V4-ARCH·Se4Op7An5Ar9Co2Ad6St8Sc3Sa5So1Br2Fo6"
)
vg_reviewer = VerigentAgent(
    reviewer,
    "VG:REVIEWER-02:V3-ANAL·Se6Op3An8Ar4Co5Ad3St7Sc6Sa7So2Br1Fo2"
)

# Create trust-aware GroupChat
group = VerigentGroupChat(
    agents=[vg_coder, vg_reviewer, user],
    task_type="code",  # prioritises Architect, Forge, Analyst scores
    max_round=12,
)

Trust-Weighted Speaker Selection

When VerigentGroupChat selects the next speaker, it:

1. Parses each agent's VG key (tier + 12 class scores)
2. Computes a relevance score based on the task_type
3. Combines tier (40%) and relevance (60%) into a composite weight
4. Selects the highest-scoring eligible agent

Task Types

VG Key Format

VG:{NAME}-{SUFFIX}:{TIER}-{PRIMARY}·{12×class_code+digit}

Example: VG:JARVIS-0A:V3-ARCH·Se4Op7An5Ar9Co2Ad6St8Sc3Sa5So1Br2Fo6

• Tier: V0 (unverified) to V6 (maximum trust)
• Primary: Dominant capability class
• Scores: 12 class scores (0-9), each representing verified capability

Transparency Log

Access the trust decision log for auditability:

for entry in group.trust_log:
    print(entry["selected"], entry["candidates"])

API Reference

VerigentAgent(agent, vg_key)

Wraps a ConversableAgent with Verigent identity. Injects the key into the system message and parses keys from incoming messages.

VerigentGroupChat(agents, task_type, trust_bias, ...)

Extends GroupChat with trust-weighted speaker selection. Set trust_bias=0 to disable trust influence, trust_bias=1 for pure trust-based selection.

parse_key(text) -> VGKey | None

Parse a VG key from any text string.

evaluate_trust(key, task_type) -> TrustScore

Compute trust score for a key given a task context.

Links

• Verigent — Agent trust verification
• AutoGen — Multi-agent framework
• PyPI

License

MIT