Gemini - Harmonia

Gemini supplies multimodal and text reasoning for most Harmonia cognitive roles. Role policy fixes the configured model, generation limits, eligible tasks, and timeout.

How Harmonia uses it

Work	Configured model family
Transcription and source understanding	Gemini 3.5 Flash, including authorized multimodal evidence
Ryan strategy	Gemini 3.5 Flash
Dara editorial assessment	Gemini 3.5 Flash
Maya presentation planning	Gemini 3.5 Flash
Nova read-only liaison	Gemini 3.5 Flash
Coordinator and Temi planning	Gemini 3.5 Flash-Lite

Inputs are typed and bounded. Source-derived claims must retain supplied evidence references. The application reserves budget before invocation, records normalized usage, and validates output before advancing a stage.

Security and authority

Gemini cannot approve strategies or effects, mutate Firestore, schedule externally, publish, or create verification receipts. Multimodal input is attached only from an authorized source package.

Failure behavior

Unavailable configured models fail visibly. Output that violates a schema, evidence boundary, or role boundary is rejected as a protocol failure.

Evidence status

Model routing and output contracts are locally tested. Authenticated Gemini calls and their billed usage require captured provider and Cloud Trace evidence.

Relevant files

agent/harmonia_agent/role_models.py
agent/harmonia_agent/model_governance.py
agent/harmonia_agent/model_catalog.py
agent/harmonia_agent/content.py
agent/harmonia_agent/agents.py
docs/models-cost-evaluation.mdx

Last modified on August 28, 2026