Escalation Policy
Defines severity levels, triggers, and a notification path for incidents involving autonomous agents, with a rollback procedure for Phase B degradation.
What this file does
Defines severity levels, triggers, and a notification path for incidents involving autonomous agents, with a rollback procedure for Phase B degradation.
When to use it
- Setting up incident response for agent-driven workflows
- Defining clear escalation triggers for credential leaks or out-of-scope actions
- Establishing a rollback plan when agent behavior degrades
- Creating a policy for when to notify a human founder directly
Escalation Policy
Status: Active | Owner:
matt@appyhourlabs.com| Version: 1.0 · 2026-02-21
Purpose
Provide a clear, fast path for handling risk, ambiguity, and incidents. When in doubt: stop, log, notify. Do not guess at what to do with a situation that isn't covered by another policy.
Severity Levels
| Severity | Definition | Examples | Response Time |
|---|---|---|---|
| SEV-1 | Production impact, data at risk, or agent acting out of scope | Credential in repo, autonomous send without approval, agent accessing human-only account | Immediately — stop everything |
| SEV-2 | Policy breach, degraded output quality, manual intervention required | Gate failure ignored, scope creep discovered, off-brand external post | Within 1 hour |
| SEV-3 | Noticeable issue, contained, no external impact | Broken internal link, failed pipeline with no consequence, minor formatting violation | Within 24 hours |
Escalate Immediately (SEV-1 Triggers)
Stop automated execution and notify matt@appyhourlabs.com if:
- A credential, API key, or token appears in any committed file, log, or agent output
- An agent takes or prepares to take any action on
legal@appyhourlabs.com,security@appyhourlabs.com, orbilling@appyhourlabs.comsystems - Any workflow initiates or proposes actual money movement
- A request is received to enable domain-wide delegation
- An external communication was sent without explicit human approval
- An agent acts materially outside its documented task scope
Escalation Path
1. Stop automated execution immediately
2. Preserve the state (don't delete logs or revert blindly)
3. Notify matt@appyhourlabs.com with: what happened, when, what was affected
4. File an incident report: INCIDENTS/ (use RUNBOOKS/incident-report-template.md)
5. Wait for explicit human direction before resuming any related work
When to Ping the Founder
Tag <@U08R1DR4LFJ> in #ai-office (see RUNBOOKS/human-escalation.md for format) for:
- All SEV-1 incidents
- Any SEV-2 incident involving external systems or communications
- Any situation where no existing policy covers the decision
- Any request to deviate from or modify a guardrail
- Any agent output that requires Matt to take action (merge, send, approve, review)
For SEV-3 situations not involving the founder's direct attention, document the issue in a task file and handle in the normal daily review cycle.
Rollback to Phase A
If Phase B is in effect and behavior degrades:
- Rotate or revoke the agent's outbound publish credentials immediately
- Update
POLICIES/agent-registry.mdto set the agent tier back tophase-a - File a SEV-1 incident report
- Conduct a blameless post-mortem within 5 business days
- All Phase A → Phase B promotion counters reset to zero
Phase A rollback can be initiated by matt@appyhourlabs.com at any time without justification required.
Related: POLICIES/ai-safety-charter.md · POLICIES/phase-a-to-b.md · RUNBOOKS/incident-report-template.md
What's inside
5 severity levels, 6 SEV-1 triggers, an escalation path, founder ping criteria, and a Phase A rollback procedure.
Change this for your project
- Replace
matt@appyhourlabs.comwith your own escalation contact - Replace
<@U08R1DR4LFJ>with your own Slack user ID - Replace
#ai-officewith your own Slack channel name - Replace
AppyHourLabs/ai-workforce-labreferences with your own repository paths
Where it goes
Keep it in your repository where the agent or team that needs it will read it.
Worth borrowing
- Stop, log, notify as a default response to ambiguity
- Preserve state before reverting during an incident
- Blameless post-mortem within 5 business days after rollback
Related Documents
OpenCode Agents
Provides a collection of specialised Markdown agent files for OpenCode, each with a distinct role and thinking algorithm.
🌟 GitHub MCP Server - Feature Showcase
Showcases eight GitHub MCP server tools with example inputs, outputs, and real-world use cases for AI agents.
Chloe 部署文档
Guides you through deploying an enhanced SillyTavern fork with direct, Docker, and reverse-proxy setups, plus security hardening and troubleshooting.
Software License Compliance Report
Documents that all dependencies in an early cancer prediction project are FOSS and OSI-approved, confirming commercial use is permitted.