AI Agents and Skills

TDCFlow uses capability planning, skills, and specialist agents to keep Copilot responses product-aware and domain-specific.

Capability Planning

Capabilities classify the user's broad intent, such as:

• Project delivery intelligence.
• Client account intelligence.
• Revenue and billing intelligence.
• Resource and capacity intelligence.
• Time and utilization intelligence.
• Expense review.
• Approval queue triage.
• Settings management.
• Workflow execution.

Capabilities define the first safety boundary: what kind of answer is expected and which tools are allowed.

Skill Planning

Skills are playbooks for specific jobs. A skill contains:

• Intent examples.
• Keywords.
• Required evidence.
• Allowed tools.
• Output contract.
• Safety rules.
• Whether it can propose actions.
• Supported proposal types.

Skills should be broader than one hard-coded command. For example, a "delivery risk review" skill can answer risk questions, summarize blockers, and draft follow-up actions.

Master Orchestrator

The master orchestrator:

• Receives the user's request.
• Uses capability and skill planning.
• Decides whether direct tools are enough.
• Delegates specialist work when a request spans a domain.
• Combines specialist results into one user-facing answer.

The master must not reveal internal delegation or agent names unless the user explicitly asks about the AI system architecture.

Specialist Agents

Specialist agents are scoped by domain.

Agent	Primary use
Project Delivery Agent	Project health, risk, tasks, milestones, closeout readiness.
Client/Account Agent	Account briefs, client prep, onboarding, handoff readiness.
Revenue & Billing Agent	Invoices, reminders, billing exposure, credit note context.
Resource & Capacity Agent	Staffing, skills, utilization, workload, allocations.
Expense Agent	Expense reports, billable spend, approval triage.
Time & Utilization Agent	Timesheets, logged hours, utilization, time gaps.
Approval Queue Agent	Pending approvals and decision readiness.
Document Agent	Status updates, summaries, client-ready drafts.
Settings Agent	Modules, settings, roles, permissions, configuration guidance.
Workflow Agent	Multi-step plans and cross-module operational workflows.
General Agent	Fallback for broad or low-risk questions.

Delegation Rules

Use specialist agents when:

• A request needs domain-specific reasoning.
• Multiple modules are involved.
• A direct answer would require combining several tool results.

Do not delegate when:

• The question is simple and answerable from current context.
• The user only asks to navigate.
• The request is unsupported and needs a clear limitation.

Tool Scoping

Each agent should receive only the tools needed for its domain. This limits accidental reasoning outside scope and reduces noisy answers.

Examples:

• Billing agent can read invoice and account billing context.
• Resource agent can read skills, roles, capacity, allocations, and tasks.
• Settings agent should guide configuration but should not invent settings changes.

Adding a New Skill

When adding a skill, define:

1. Intent examples.
2. Keywords.
3. Required evidence.
4. Allowed tools.
5. Output contract.
6. Safety rules.
7. Proposal types, if applicable.

Then test:

• A read-only question.
• A request with missing data.
• A data-changing request.
• A prompt that should not trigger the skill.

Related Docs

• AI System Overview
• AI Tooling and Context
• AI Safety and Answer Rules