---
title: Claude Managed Agents Governance
description: Cromus is the pre-execution governance layer for Anthropic's Claude Managed Agents (managed-agents-2026-04-01 beta). Compile SOPs into SKILL.md, validate against ETHOS.md, simulate cost, and emit a ready-to-deploy Claude Managed Agent JSON configuration.
canonical: https://cromus.ai/claude-managed-agents
source_html: https://cromus.ai/claude-managed-agents
---

# Claude Managed Agents Governance

> Pre-execution governance for Anthropic's Claude Managed Agents runtime.

Anthropic's Claude Managed Agents (managed-agents-2026-04-01 beta) is a managed agent runtime that runs Claude as an autonomous agent in cloud containers with bash, file operations, web search, web fetch, and MCP server tools. It is a powerful runtime — but it does not answer the upstream question of *what the agent should do, what it should cost, or which policy it should obey*.

Cromus answers those questions before the agent ever spins up.

---

## What Cromus produces for Claude Managed Agents

For every workflow you bring to Cromus, the platform emits:

1. **A compiled SKILL.md** — the capability spec, with model tier (Claude Opus / Sonnet / Haiku), token budget, IO contract, and fallback policy
2. **A validated ETHOS.md** — the behavioural policy: escalation rules, refusal conditions, transparency requirements
3. **A cost simulation** — per-run and monthly projections across the Claude model family, with parallelisation savings and Croms™ waste score
4. **A Claude Managed Agent JSON configuration** — ready to POST to Anthropic's Agent endpoint, including model selection, tool list, system prompt derived from SKILL.md, and ETHOS-driven guardrails

All artefacts are produced deterministically with zero LLM calls.

---

## Why Anthropic users need pre-execution governance

Claude Managed Agents excels at execution but assumes you've already decided:

- Which Claude model to use (Opus / Sonnet / Haiku) and why
- What budget the agent is allowed to consume
- Which tools the agent may invoke and under what conditions
- How the agent should escalate, refuse, or hand off
- What the workflow's structural waste profile looks like

Cromus produces all of these as portable artefacts before the agent runs. The same artefacts work across other runtimes (Google AI Studio, Bedrock, generic MCP), so workflow governance is not Anthropic-locked.

---

## Integration paths

- **Web app** — design the workflow in Skill Studio, export the Claude Managed Agent JSON config
- **MCP server** — call `score_skill`, `validate_skill`, `validate_ethos`, `simulate_cost` from any MCP-compatible client (including Claude Cowork)
- **Cowork connector** — direct HTTP MCP at `https://mcp.cromus.ai/mcp`, six deterministic tools available inside any Cowork session
- **Skill packs** — pre-built Claude-ready workflow templates available at `/app/skills/samples` and `/ai-skill-packs`

---

## What this is not

- **Not a Claude proxy** — Cromus never sees a production token
- **Not a runtime competitor** — Anthropic still runs the agent; Cromus governs the inputs
- **Not Anthropic-exclusive** — the same SKILL.md / ETHOS.md works on Google AI Studio, Bedrock, LangGraph, CrewAI, AutoGen

---

## Related

- [Pre-Execution Governance](/pre-execution-governance.md)
- [Claude Cowork MCP Connector](/cowork-connector)
- [Google AI Studio (Gemini Interactions) Governance](/google-ai-studio)
- [SKILL.md Validator](/validator/skill.md)
- [ETHOS.md Validator](/validator/ethos.md)