---
title: Pre-Execution Governance — The upstream layer for AI agent stacks
description: Cromus is the pre-execution governance layer that scores, validates, and shapes AI workflows before tokens are spent — upstream of runtime agent stacks like Anthropic, Google Agent Platform, AWS Bedrock Agents, and Azure agent gateways.
canonical: https://cromus.ai/pre-execution-governance
source_html: https://cromus.ai/pre-execution-governance
---

# Pre-Execution Governance

> The layer of agent governance that scores and shapes AI workflows **before tokens are spent**.

Cromus sits **upstream** of runtime agent platforms. Runtime stacks (Anthropic Claude Managed Agents, Google Agent Platform, AWS Bedrock Agents, Azure agent gateways) secure and execute agents that are already deployed. Cromus governs them before they ever run.

---

## Why pre-execution matters

Most AI governance tools operate after a workflow is already running — they observe traces, flag policy violations, and meter spend in retrospect. By then, the cost has already been incurred and the wrong model has already been chosen.

Pre-execution governance asks a different question: **should this workflow run at all, in this shape, on this model?** It answers it deterministically, with zero LLM calls, before any token is spent.

---

## The three layers of agent governance

| Layer | Question it answers | Where Cromus operates |
|-------|---------------------|------------------------|
| **Layer 1 — Compile-time** | Is the workflow structurally sound? Does it match policy? | ✅ Score SOP, validate SKILL.md, validate ETHOS.md |
| **Layer 2 — Pre-execution** | What will this cost, in dollars, latency, and risk? Which model fits? | ✅ Cost simulator, model recommendation, Croms™ waste score |
| **Layer 3 — Platform governance** | Across an entire workspace, what budget, what tier, what model channels are allowed? | ✅ Workspace scope, ecosystem-tier × cost-mode matrix, model gates |
| Layer 4 — Runtime | Is the deployed agent behaving safely right now? | Anthropic, Google Agent Platform, AWS Bedrock Agents, Azure |
| Layer 5 — Post-execution | What actually happened? What did it cost? | Helicone, Langfuse, LangSmith, observability tools |

Cromus owns layers 1–3. It does **not** compete with runtime stacks — it produces the governed input that runtime stacks execute.

---

## What pre-execution governance produces

Every workflow that passes through Cromus emits a portable, deterministic artifact:

- **SKILL.md** — the capability spec (model tier, token budget, IO contract, fallback policy)
- **ETHOS.md** — the behavioral policy (escalation rules, transparency, refusal conditions)
- **MEMORY.md** — the memory contract (what is retained, compressed, transferred)
- **Cost simulation** — per-run + monthly projections across 5 cost-quality modes and 62 verified models
- **Croms™ score** — quantified preventable waste (cost, latency, failure, structural)
- **Platform-specific config** — Claude Managed Agent JSON, Gemini Interactions API config, generic MCP, LangGraph/CrewAI/n8n/AutoGen exports

All of this is generated **before** a single token is spent.

---

## Workspace scope and the ecosystem-tier × cost-mode matrix

Pre-execution governance at the platform level is described along two orthogonal axes:

- **Cost-quality mode** — Eco, Cost, Balanced, Quality, Open Source. Answers *"how much performance do I need?"*
- **Ecosystem tier** — Mainstream, Managed, Aggregator, Infra. Answers *"how do I reach the model?"*

A workspace's scope (enabled channels + optional explicit tier preferences) gates which models the recommendation engine may suggest. Every model surfaced anywhere in the product carries an **EcosystemBadge** tier chip plus a setup-effort hint, so governance decisions are legible at a glance.

---

## What Cromus is *not*

- Not an LLM observability tool — it never sees a production token.
- Not a runtime safety stack — it doesn't intercept calls at execution time.
- Not a model gateway — it doesn't proxy traffic.
- Not a competitor to LangGraph, CrewAI, n8n, AutoGen — it compiles to all of them.

Cromus is the **upstream governance + scoring layer** that decides whether a workflow is fit to run, what it should cost, and which runtime should execute it.

---

## Related concepts

- [Croms™ — preventable AI workflow waste](/croms.md)
- [Workflow Intelligence as a Service](/workflow-intelligence-as-a-service.md)
- [Claude Managed Agents Governance](/claude-managed-agents.md)
- [SKILL.md Validator](/validator/skill.md)
- [ETHOS.md Validator](/validator/ethos.md)
- [MEMORY.md Validator](/validator/memory.md)