Module 8: AI Assistants, Agents, Connectors, and Automation

Learning Objectives

  • Distinguish chat, assistants, workflows, automations, agents, and multi-agent systems.
  • Use files and connected sources as controlled context.
  • Design permissions and approval gates for tool-enabled workflows.
  • Identify high-value teaching, research, and administrative agent use cases.
  • Design a simple teaching assistant or bounded agent without programming expertise.
  • Understand when orchestration adds value and when it creates unnecessary complexity.

Connector

An authorized integration allowing an AI system to access or act on information in another application within defined permissions.

Tool Use

The ability of an AI system to call a defined capability such as search, file retrieval, calculation, or an authorized action.

Automation

A predefined process that executes a task or sequence with limited real-time human instruction.

Orchestration

Coordination of multiple steps, tools, agents, or specialized roles toward a shared objective.

Approval Gate

A required human authorization point before a consequential action or workflow stage proceeds.

When to Move Beyond Chat

A normal chat is often enough for one-off work. Move toward assistants, workflows, automation, or agents only when the task benefits from persistent instructions, repeated context, multiple tools, multi-stage state, or scheduled/conditional execution.

Ask four questions: Does the task repeat? Does it require authoritative files or connected systems? Does it involve multiple dependent stages? Can the next action be safely bounded and verified?

If the answer is mostly no, a prompt is simpler and easier to audit. If several answers are yes, a workflow may provide value. Agentic systems should earn their complexity through reduced repetition, stronger traceability, or better coordination—not novelty.

Files and Knowledge Sources

A course assistant may use a syllabus, approved notes, instructor FAQ, policy, readings, or other controlled sources. Define which materials are authoritative and how conflicts are handled. Avoid dumping an entire drive into context when only two files matter.

For persistent knowledge, maintain revision identity and removal procedures. If an outdated file remains in the knowledge set, an agent can repeatedly reproduce obsolete guidance.

Require source traceability. A teaching assistant that explains a course policy should identify the file or authoritative rule behind the answer. If no source supports the claim, it should flag the gap rather than improvise.

Connectors and Tool Use

Connectors extend an AI system into other applications or data sources. Examples include documents, cloud drives, spreadsheets, calendars, email, project systems, or specialized tools. Availability depends on the platform and organizational configuration.

Distinguish read from write access. Reading a syllabus is lower consequence than editing it. Drafting an announcement is different from sending it. An analytics workflow can read a de-identified spreadsheet without being allowed to modify authoritative grades.

Before connecting a tool, document the purpose, data available, actions allowed, actions prohibited, approval required, and audit evidence. This permission contract should remain independent of model capability.

Least Privilege and Approval Gates

Use least privilege: grant only the access necessary for the current task. Separate reversible drafting actions from consequential external actions such as sending, publishing, deleting, changing grades, or updating official records.

A practical permission ladder is Read → Draft → Edit Controlled Artifact → Send/Publish → Delete/Irreversible Action. The farther right the workflow moves, the stronger the approval requirement should become. Consequential steps should pause for explicit human review with a clear summary of what will happen.

Also define stop conditions for missing evidence, authentication failure, conflicting sources, unexpected tool responses, or a request outside the agent's scope.

Designing a Simple Teaching Agent

A simple teaching agent needs more than a persona. Specify a bounded objective, approved sources, tools and permissions, state it may retain, step sequence or planning policy, output schema, stop conditions, evaluation cases, and human approval.

For example, a course assistant can answer from approved materials, create practice questions, and draft study guidance while being prohibited from changing grades, inventing policy, or contacting learners automatically.

Keep the architecture observable. The educator should be able to determine what evidence was used and what action the agent proposes.

Orchestration and Multi-Agent Work

Multi-agent systems are useful when distinct roles genuinely benefit from separation. A teaching-material workflow might use Planner → Creator → Reviewer → Verifier → Human.

Each role should receive defined inputs and produce a structured handoff. The reviewer should not merely restate the creator's answer; it should inspect against explicit criteria. The verifier should check source support, calculations, or policy as appropriate.

Avoid orchestration theater—five agents all paraphrasing the same context. Decomposition should reduce conflict, improve specialization, or create independent checks. More agents increase coordination cost, context transfer, and failure surfaces.

Administrative and Research Workflows

Faculty workflows can include literature triage, course-material organization, announcements, meeting summaries, calendar preparation, accreditation evidence gathering, CQI documentation, or draft reports. These are often good agentic candidates because they contain repeatable steps and structured evidence.

The same controls still apply. An assistant may draft an announcement from an approved calendar, but sending it should require review. A research assistant may organize literature, but citations must be verified. A documentation agent may assemble evidence, but it should not claim that a requirement is satisfied unless the evidence actually demonstrates it.

Evaluation, Failure Testing, and Course Delivery

Test the workflow before trusting it. Include normal cases, missing files, conflicting instructions, tool failure, unauthorized action attempts, and adversarial content such as a document that tries to override system instructions.

Record whether the agent stops safely. Tool outputs and external content should be treated as untrusted input until validated.

Use 90–110 minutes for the full module or 60–70 minutes condensed. The hands-on activity should create a real agent specification even if the participant uses an existing assistant rather than writing code.

AI Assistants, Agents, Connectors, and Automation — Applied Workflow

  1. Define the faculty or teaching objective and boundary of responsibility.
  2. Identify approved files, data sources, connectors, and tools.
  3. Assign least-privilege permissions to each tool or stage.
  4. Define agent tasks, stop conditions, and required outputs.
  5. Insert human approval before consequential actions or external release.
  6. Test normal, missing-data, tool-failure, and unauthorized-action cases.
Agentic Teaching Workflow

Coordinate evidence, tools, agent tasks, verification, and human approval for a bounded teaching objective.

Agentic Teaching WorkflowCoordinate evidence, tools, agent tasks, verification, and human approval for a bounded teaching objective.. Define teaching objective and success criteria → Authorize files, knowledge sources, tools, and permissions; Authorize files, knowledge sources, tools, and permissions → Execute bounded planning, drafting, or analysis steps; Execute bounded planning, drafting, or analysis steps → Check evidence, tool results, and policy boundaries; Check evidence, tool results, and policy boundaries → Does the next step create a consequential or external action?; Does the next step create a consequential or external action? — No → Store or release the reviewed result; Does the next step create a consequential or external action? — Yes → Require explicit human approval; Require explicit human approval → Store or release the reviewed result

Define teaching objective and success criteria → Authorize files, knowledge sources, tools, and permissions; Authorize files, knowledge sources, tools, and permissions → Execute bounded planning, drafting, or analysis steps; Execute bounded planning, drafting, or analysis steps → Check evidence, tool results, and policy boundaries; Check evidence, tool results, and policy boundaries → Does the next step create a consequential or external action?; Does the next step create a consequential or external action? — No → Store or release the reviewed result; Does the next step create a consequential or external action? — Yes → Require explicit human approval; Require explicit human approval → Store or release the reviewed result

  • Define teaching objective and success criteria: terminator
  • Authorize files, knowledge sources, tools, and permissions: process
  • Execute bounded planning, drafting, or analysis steps: process
  • Check evidence, tool results, and policy boundaries: process
  • Does the next step create a consequential or external action?: decision
  • Require explicit human approval: process
  • Store or release the reviewed result: terminator
Chatbot to Agent Architecture

Follow the sequence from prompt to context-aware assistant, tool-enabled workflow, bounded agent, and orchestration. Each step adds state or action capability and therefore additional control requirements.

Prompted chat
Context-aware assistant
Tool-enabled workflow
Bounded agent
Orchestrated system
Tool and Connector Permission Stack

Read the stack as a permission ladder. Read access is the baseline; draft, edit, publish, and irreversible actions require progressively stronger controls and approval.

1
Read
2
Create draft
3
Edit controlled artifact
4
Send or publish
5
Delete or irreversible action
Planner Creator Reviewer Human

Use the sequence as an orchestration pattern: planner defines work, creator produces, reviewer challenges, verifier checks evidence, and the human resolves consequential decisions.

Planner
Creator
Reviewer
Verifier
Human approval
Automation Complexity Ladder

The grid is a complexity test. Prefer a simple prompt or repeatable workflow unless tool use, state, or autonomous continuation creates measurable value.

Simple prompt
Repeatable workflow
Tool-enabled automation
Agent only when needed

Permissions Are Separate from Model Intelligence

A capable model should still have only the access and actions required for the task.

Do Not Automate Consequential Actions by Default

Sending, publishing, deleting, grading, or changing authoritative records should require explicit controls and appropriate approval.

Key Takeaways
  • Agentic engineering is controlled workflow design, not simply longer prompting.
  • Files and connectors should be bounded sources of evidence and action.
  • Least privilege reduces risk and review burden.
  • Simple teaching agents are useful when tasks are repetitive and evidence-rich.
  • Orchestration should earn its complexity through measurable value.

References