Policies
What a policy says, and how to move it from audit to enforce.
What a policy says
A policy is a YAML file in your agents.d directory, validated against the published schema (agentgarm.com/schemas/agents/v1). It names the agent it covers, its mode, how it fails, and the boundaries the agent works within. The opening of the default Claude Code pack:
# AgentGarm default policy pack: Claude Code.
# Install with `agentgarm policy init claude-code`. Packs ship in audit mode: run
# `agentgarm policy lint --blast-radius` on your own history before promoting to enforce.
schema: agentgarm.com/schemas/agents/v1
name: claude-code
version: "1.1.0"
mode: audit
profile: interactive
on_internal_error: fail_open_alert
match:
launcher: true
binaries: [claude]
process_tree_root: true
swarm_governance:
max_subagent_depth: 4
max_concurrent_subagents: 8
loop_detection:
max_identical_tool_calls: 8
window_seconds: 120
action: alert
security_boundaries:
filesystem:
# Reads of credentials are denied; .env.example and other templates stay readable.
deny:
- "~/.ssh"
- "~/.aws"
- "~/.gnupg"
- "~/.config/gcloud"
- "~/.azure"
- "~/.kube/config"
- "~/.netrc"
- "~/.git-credentials"
- "~/.docker/config.json"
- "~/.config/gh/hosts.yml"
- "**/.env"
- "**/.env.local"
- "**/.env.*.local"
# Configuration that runs commands outside any sandbox: file tools may not change it.Audit, then enforce
Packs ship in audit mode: every decision is made and recorded, and nothing is blocked. Check a candidate against what your agents actually did before you enforce it:
agentgarm policy lint ~/.config/agentgarm/agents.d/claude-code.yaml --blast-radiusIt reports shadowed and conflicting rules, and what the policy would have denied in your recent history.
Fail modes
Each boundary declares how it fails: fail_closed denies when AgentGarm cannot decide, and fail_open_alert lets the action through and raises an alert. Where each is enforced.