| name | agent-safety |
| description | Behavioral constraints for Make.com autonomous agents — what may and may not be done without explicit human confirmation during /factory and /make sessions |
| allowed-tools | Read, Glob, Grep |
Agent Safety — Autonomous Operation Constraints
Purpose
Define the boundary between actions an agent may take autonomously and actions
that require an explicit human confirmation step. These rules apply to ALL agents
in this plugin, especially scenario-orchestrator during /factory runs.
NEVER Without Human Confirmation
These actions require the approval block (see CLAUDE.md ## Approval Gates) and
an explicit "approve" reply before any MCP call is made:
| Action | Reason |
|---|
scenarios_activate | Starts consuming operations and billing |
scenarios_delete | Permanent — cannot be recovered |
hooks_delete | Breaks all upstream triggers immediately |
data-stores_delete | All records lost permanently |
data-store-records_delete (bulk) | Mass data loss |
| Any production scenario modification | Risk of breaking live automation |
| Third-party API call that incurs cost | Credits consumed per call |
| Adding a new Make.com connection | Requires OAuth from user |
MAY Do Autonomously (No Confirmation Needed)
| Action | Notes |
|---|
scenarios_list, scenarios_get | Read-only, no side effects |
scenarios_create in draft state | Not active, no operations consumed |
scenarios_update on a draft | Safe — scenario not yet active |
validate_blueprint_schema | Validation only |
app_documentation_get | Documentation fetch only |
Write to .make/ directories | Local audit trail only |
Read any .make/ file | Memory and context reads |
Graduated Trust During /factory
The orchestrator runs four phases. Trust escalates with each approval:
- Kickstart — All reads are autonomous. No writes.
- Bootstrap — Writes
.make/context/ files. Autonomous.
- System Design — Writes
.make/plans/. Shows plan. STOP for approval.
- Sprint — Executes plan one step at a time. Each MCP write requires approval.
Never batch approvals. Each write MCP call gets its own confirmation.
On Unexpected State
If the agent encounters unexpected state (scenario already exists, webhook URL changed,
connection expired), it must:
- Stop immediately — do not attempt to work around it
- Surface the discrepancy to the user in plain language
- Propose options and wait for a decision
Never silently compensate for unexpected state.