The AI COO and agentic workflows
The AI COO is InBlack’s tenant-aware agent experience. Portal provides the Conversation Theatre and briefing UI; the agent discovers and calls the same MCP catalog available to authorized external clients.
What controls an agent
Tool availability is the intersection of several gates:
- Catalog surface (
publicorinternal) and tool risk tier. - Billing pillar and AI COO scope (
accounting,full, orunlimited). - Required feature flags and entitlements declared by the tool.
- Authenticated principal, tenant membership, role, and permission checks.
- Per-tenant agent policy and model-policy decisions at execution time.
Passing one gate never bypasses the others. A tool visible in the generated catalog may still be unavailable to a specific tenant or caller.
Reads, previews, and writes
The current catalog contains read, write, and generated simulate tools. A simulate companion previews its paired catalog write without committing that write. Write tools remain subject to their own confirmation, pending-action, authorization, and audit behavior; the tier alone is not a universal approval state machine.
Customer-facing financial writes remain MCP-first. Product UIs continue to call product APIs, and code-mode/API use follows the ADR-0019 carve-outs documented in MCP vs. HTTP APIs.
Where work appears
- Portal
/hosts the Conversation Theatre and briefing card states. - Accounting
/ai-actionshosts pending AI actions. - Accounting
/agent-activityshows the activity timeline. - Accounting
/agent-alerts-settingsmanages agent alert settings.
These are separate product views over agent and audit data; the guide does not assume every MCP call appears identically in every view.
For the current tool contract, use The MCP tool catalog, Pending actions, and Tool errors.