Skip to content

ANTHROPIC-INTEGRATION-PROPOSAL.md

Latest commit

 

History

History
108 lines (84 loc) · 4.22 KB

ANTHROPIC-INTEGRATION-PROPOSAL.md

File metadata and controls

108 lines (84 loc) · 4.22 KB

Anthropic — Governance Integrations

Submissions:

Overview

Three complementary submissions to the Anthropic ecosystem providing governance patterns for Claude-based agent systems. Each targets a different integration point.

1. Agent Governance Skill (PR #424)

Status: Open PR — awaiting review Closes: anthropics/skills#412 (invited by @BennyTheBuilder)

What's Included

Pattern Description
Governance Policy Model Deny-by-default, tri-state decisions (allow/deny/review), YAML-based
Policy Composition Org → Team → Agent layering, stricter always wins
Threat Detection 5 categories: data exfiltration, prompt injection, privilege escalation, credential harvesting, destructive operations
Trust Scoring Decay-based scores with asymmetric reward/penalty for multi-agent delegation
Audit Trail Append-only, tamper-evident SHA-256 Merkle chain
Tool-Level Governance Single decorator adds policy enforcement + audit to any tool function
Framework Integration Examples for PydanticAI, CrewAI, OpenAI Agents SDK, Google ADK

Design Decisions (per @BennyTheBuilder feedback)

  • Leads with the policy model (not frameworks or decorators)
  • Includes a minimal working example (copy-paste YAML + Python)
  • Audit trail design is front-and-center (Merkle chain with what/who/when/why)
  • Fail-closed everywhere — any evaluation error returns deny

2. Claude Plugin Proposal (Issue #415)

Status: Open issue — concept validation stage

Plugin Structure

agent-governance/
├── .claude-plugin/plugin.json
├── .mcp.json                    # Governance MCP server
├── commands/
│   ├── governance-check.md      # /governance-check — audit current project
│   └── policy-init.md           # /policy-init — scaffold governance config
├── agents/
│   └── governance-reviewer.md   # Agent that reviews code for governance
├── skills/
│   └── agent-governance/SKILL.md
└── README.md

MCP Server Tools

  • governance/check_policy — Validate tool calls against policies
  • governance/detect_threats — Scan code/prompts for threat patterns
  • governance/audit_query — Query audit trail for compliance

3. Governance Cookbook (Issue #384)

Status: Open issue — awaiting approval

Proposed Notebook: misc/agent_governance.ipynb

Sections:

  1. Introduction: Why Agent Governance Matters
  2. Defining a Governance Policy (YAML-based, composable)
  3. Tool Policy Enforcement (allowlists, blocklists, argument validation)
  4. Threat Detection Patterns (5 categories with Claude-based classification)
  5. Trust Scoring for Multi-Agent Delegation
  6. Building Append-Only Audit Trails
  7. Putting It Together: A Governed Agent Pipeline

Code Preview

import anthropic

async def governed_tool_use(client, policy, messages, tools):
    response = await client.messages.create(
        model="claude-sonnet-4-20250514",
        messages=messages, tools=tools,
    )
    for block in response.content:
        if block.type == "tool_use":
            if block.name in policy.blocked_tools:
                raise GovernanceViolation(f"Tool '{block.name}' blocked")
            if policy.threat_detection:
                threats = detect_threats(block.input)
                if threats:
                    raise GovernanceViolation(f"Threat: {threats}")
    return response

Links