Operations

AI Agents

Phase 10 Safe Foundation. Mock-only runtime, human approval for every side effect.

Phase 10 — AI Agent Operating Layer (Safe Foundation). Thirteen new tables: agents, prompt versions, tool registry, agent/tool permissions, conversations, messages with provenance, tool-call logs, an action approval queue, guardrail events, and eval sets/cases/runs/results. Live Claude is disabled (ANTHROPIC_AI_ENABLED=false). All output is mock-generated and clearly labelled. Sensitive tools (payout approval, commission rule changes, communication send, policy binding, final compliance bundles) are forbidden in Phase 10. Every side effect requires a human reviewer in the approval queue.

Active agents

4 draft

Enabled tools

23 registered

Sensitive tools (disabled)

Forbidden in Phase 10

Pending approvals

1 guardrail events

Agents

Role-aware agents. Active agents are visible in their scoped dashboards. Drafts are not yet wired to any surface.

Managing Director

Earnest — Managing Director Agent

Active

Strategic copilot for the Managing Director: pipeline summaries, financial readiness, executive briefings.

Prompts: 2
Tools: 0/0
Convos: 0

Office Manager

Donna — Office Manager Agent

Active

Operational copilot for the Office Manager: daily task triage, commission discrepancy review, calendar oversight.

Prompts: 2
Tools: 0/0
Convos: 1

Staff Copilot

Active

General staff copilot scoped to the user's permissions. Read/draft/propose only.

Prompts: 1
Tools: 0/0
Convos: 0

Website Agent

Website Knowledge & Scheduling Agent

Draft

Public-surface agent for eapinder.com FAQs and appointment intent. Cannot read internal client data.

Prompts: 1
Tools: 0/0
Convos: 0

Compliance

Compliance Agent

Draft

Reads compliance readiness, drafts compliance checklists. Cannot generate final compliance bundles.

Prompts: 1
Tools: 0/0
Convos: 0

Event

Event Agent

Draft

Drafts event recaps, suggests follow-ups, summarises attendance. Read/draft/propose only.

Prompts: 1
Tools: 0/0
Convos: 0

Commission

Commission / Reconciliation Agent

Draft

Drafts commission discrepancy resolutions for Donna. Cannot approve payouts or change commission rules.

Prompts: 1
Tools: 0/0
Convos: 0

AI Readiness

Phase 10 — Safe Foundation

ANTHROPIC_AI_ENABLED
Live Claude calls are disabled. All Phase 10 AI output is mock-only.
disabled
Live tool execution
All tool calls run in mock execution mode.
disabled
Sensitive tools
approve_payout_batch, change_commission_rule, send_communication, bind_policy, generate_final_compliance_bundle — all forbidden.
disabled
Human approval required
Every side effect (level ≥ low) requires a human reviewer in the approval queue.
enabled
RLS enforced
All ai_* tables have row-level security forced on.
enabled

Approval Queue (mock)

AI-proposed actions awaiting a human reviewer.

Phase 10 — Mock-only. Approving / rejecting items is disabled. Execution lands in a later sub-phase.

Action	Agent	Side effect	Status
Resolve discrepancy: chargeback variance > $50 (Apr 2026) AI proposes setting the discrepancy to "needs_owner_review" and notifying the producer. Human approval required.	Donna — Office Manager Agent	Medium	Pending

Recent Guardrail Events

Refusals, blocked tool requests, and prompt-injection detections.

Title	Type	Severity	Agent	Resolution
Blocked: approve_payout_batch requested User asked the agent to approve the Apr 2026 payout batch. Tool is disabled in Phase 10. Refused.	Blocked Tool	High	Donna — Office Manager Agent	Refused. Approve payouts manually under /commissions/payouts.