---
title: Google AI Studio (Gemini Interactions) Governance — Cromus
description: Cromus compiles SOPs into SKILL.md, validates against ETHOS.md, simulates cost across the Gemini model family, and produces a Gemini Interactions API configuration for the typed steps format announced in May 2026.
canonical: https://cromus.ai/google-ai-studio
source_html: https://cromus.ai/google-ai-studio
---

# Google AI Studio — Gemini Interactions Governance

> Pre-execution governance for Google's Gemini Interactions API.

Google announced the **Gemini Interactions API** (May 2026) — every agent action (thought, function_call, function_call_result, google_search_call, google_search_result, model_output) is now its own typed step instead of strict user/model turns. Cromus provides pre-execution governance for this new agent runtime.

---

## What Cromus produces for Gemini

From a compiled SOP, Cromus generates:

1. **SKILL.md** — portable workflow capability spec with token estimates, IO contracts, and Gemini-specific configuration
2. **ETHOS.md validation** — behavioral governance check against declared policies before deployment
3. **Gemini cost simulation** — per-run and monthly cost projections across the Gemini model family
4. **Interactions API configuration object** — typed steps-response template ready for the Gemini Interactions API endpoint
5. **Croms™ score** — quantified preventable waste across cost, latency, failure risk, and structural inefficiency

---

## Gemini model coverage

Cromus's verified model registry covers the full Gemini family with weekly pricing and deprecation checks:

| Model | Tier | Typical use |
|-------|------|-------------|
| gemini-2.5-flash | Balanced | General-purpose, cost-efficient |
| gemini-2.5-pro | Quality | Advanced reasoning, complex tasks |
| gemini-2.0-flash-lite | Lightweight | High-volume, low-cost routing |
| gemini-ultra | Frontier | Maximum quality, research tasks |

---

## The Gemini Interactions API format

The Interactions API represents each agent action as a typed step object. Cromus compiles a SKILL.md into a steps-response template:

```json
{
  "steps": [
    { "type": "thought", "content": "Analyze the support ticket" },
    { "type": "function_call", "name": "classify_intent", "args": {...} },
    { "type": "function_call_result", "name": "classify_intent", "result": {...} },
    { "type": "model_output", "content": "..." }
  ]
}
```

This template is generated deterministically from the compiled SKILL.md — no LLM calls, same input always produces the same output.

---

## Pre-execution governance for Gemini

Cromus is upstream of the Gemini runtime. It produces the governed input that Google Agent Platform executes. It does not compete with Google Agent Platform — it is the layer that ensures workflows are fit to run before they reach the runtime.

---

## Related pages

- [Claude Managed Agents Governance →](/claude-managed-agents)
- [Pre-Execution Governance →](/pre-execution-governance)
- [Cowork Connector →](/cowork-connector)
- [Workflow Intelligence as a Service →](/workflow-intelligence-as-a-service)