Day 2 · Capstone · Lesson 12

Deploying Your Agentic Workflow

Course Target: Bring the whole course together — design, draft, and present one end-to-end AI automation pipeline that takes a real academic-administration bottleneck from raw input to a finished, accessible, governed output you could actually deploy.

This is the capstone. You won’t learn a new tool today; you’ll integrate what you already have — scoped research (L8), structured extraction (L9), tool connection and sandboxed analysis (L10), and the governance guardrails (L11) — into one pipeline, then strengthen it through peer optimization. Deliverable: a pipeline plan, a working draft, and a 5-minute walkthrough.

Start from your bottleneck

A concrete goal for todayPick the recurring administrative task that eats your week — the report you re-build, the data you re-type, the packet you re-assemble. By the end you’ll have a designed, partly-built, and peer-reviewed pipeline to take it off your plate — safely.

Learning Objectives

Design an end-to-end pipeline for a real academic-administration bottleneck, integrating skills from across the course (research, extraction, tool connection, analysis).
Build a working draft and add the governance guardrails (de-identification, verification, human sign-off, accessibility) so it’s safe to deploy.
Present your pipeline and give and use structured peer optimization to strengthen it before deployment.

A note on what "done" means todayYou’ll leave with a deployable draft and plan plus peer feedback — a strong first version, not a finished, fully-deployed system.

Key Terms

Select any card to flip it. Cards are keyboard-operable — press Enter or Space.

Course Outline & Lesson Modules

If you remember one thing: A deployable pipeline isn’t one clever prompt — it’s small checkable steps, a guardrail at every risky one, and a human at the irreversible step.

What you’ll deliver

A pipeline plan — the bottleneck, the outcome, and the mapped stages.
A working draft — at least the riskiest steps drafted/prompted, with guardrails attached.
A 5-minute walkthrough — what it does, how it’s guarded, and where you’re unsure — plus peer review.

What makes a pipeline deployable

Small, checkable steps — so you can verify each, not one opaque mega-step.
Guardrails — de-identification, verification, human sign-off, accessibility.
An audit trail — inputs, prompts, outputs, dates — so it’s defensible.
A clear stop/failure behavior — it asks for help or halts rather than guessing.

You’ll be assessed on the capstone rubric below — design it against those criteria from the start.

Quick check: your pipeline auto-emails a report the moment the AI finishes. What’s missing?

The human sign-off — and probably verification and an audit trail. Insert your approval before anything is sent; never let the last step be irreversible and unattended.

If you remember one thing: Every stage of a good pipeline maps to something you already learned — and to a guardrail. The capstone is assembly, not invention.

Exemplar — an accessible weekly program digest

The bottleneck: every week you hand-compile program updates and student-success numbers into a post. Here’s the same job as a guarded pipeline:

Pick a bottleneck → define the outcome
   |
   ▼  Gather / Research   (L8: a scoped loop — sources + stop condition)
   ▼  Extract             (L9: structured output, the "not found" rule)
   ▼  Connect / Analyze   (L10: approved tool, sandbox, de-identified data)
   ▼  Assemble            (data + template → draft)
   ▼  Verify              (recompute a figure / check against the source)
   ▼  Sign off            (YOU approve — human in the loop, L11)
   ▼  Deliver             (accessible output — WCAG check, L11)  + audit trail

Notice how each stage draws on a lesson and carries a guardrail: the research loop is scoped (L8); extraction asks for structured output (L9); the metrics are de-identified and analyzed in a sandbox (L10); a human signs off and the post is WCAG-checked before publishing (L11) — with an audit trail throughout.

Your turnYour pipeline won’t have all eight stages — keep only the ones your bottleneck needs. The non-negotiables are verification and human sign-off wherever the output is consequential.

Quick check: which two stages should almost never be dropped?

Verify and Sign off. Speed is the point of the others; trust and safety come from those two.

If you remember one thing: You already built the parts in Lessons 8–11. This checklist is how you make sure none of them is missing from your pipeline.

Run every item against your design — copy and keep it:

Capstone integration checklist — apply each to your pipeline:
[ ] Bottleneck & outcome — a real recurring task, with a clear definition of "done."
[ ] Scope (L8) — bounded steps and a stop condition; no endless crawl.
[ ] Extraction (L9) — ask for structured output; mark missing fields "not found."
[ ] Tools & data (L10) — only approved connections; sandbox for code; de-identify student data.
[ ] Verification — recompute a figure / check output against the source.
[ ] Human sign-off (L11) — a person approves before any irreversible action.
[ ] Accessibility (L11) — run the four-principle WCAG check on the output.
[ ] Audit trail — keep inputs, prompts, outputs, and dates.

Each item links back to a lesson’s reusable artifact: the discovery-loop instruction (L8), the extraction template (L9), the "Should I connect this tool?" decision aid and sandbox settings (L10), and the FERPA decision aid, injection defense, and WCAG pre-publish checklist (L11).

Quick check: your pipeline never touches student data. Which items still apply?

Most of them — scope, extraction quality, verification, human sign-off, accessibility of the output, and an audit trail all still matter. De-identification is the one you can relax (because there’s no PII to protect).

Capstone Rubric

Your pipeline and walkthrough are assessed on these five criteria. Design to meet them.

What "meets" looks like
Criterion	Meets when…
Integration	Uses at least three course skills in one connected flow (e.g., scoped research + extraction + analysis/connection).
Guardrails	De-identifies sensitive data, includes a verification step, and requires human sign-off before any irreversible action; output is accessible.
Reliability	Small, checkable steps; a clear stop/failure behavior; an audit trail.
Fit	Targets a real, recurring bottleneck with a clearly defined outcome.
Presentation & peer response	Clear 5-minute walkthrough that names its own risks and folds in at least one peer optimization.

Before You Present

A quick self-check: can you say, in one sentence each, (1) the bottleneck and outcome, (2) the stages, (3) where verification and sign-off sit, and (4) the one risk you most want a peer to pressure-test? If yes, you’re ready.

Hands-on Workspace Modules

Activity 01

Scope & Map

Choose your bottleneck, define the outcome, and map the pipeline stages — keeping only what you need.

real bottleneck + clear outcome;
stages mapped;
data touched / de-identification noted.

Open Activity 1

Activity 02

Build a Draft & Guardrail It

For each stage, note the tool/prompt you’d use and attach the guardrail from the integration checklist.

tool/prompt per stage;
a guardrail on every risky step;
verification + sign-off present.

Open Activity 2

Activity 03

Present & Peer-Optimize

Prepare your 5-minute walkthrough, then give a structured peer review against the rubric — and fold in what you receive.

walkthrough names its own risks;
peer review: strength + risk + optimization;
one peer idea adopted.

Open Activity 3

Wrap-up

Reflect (2 minutes)What is the single guardrail you’ll never skip when you deploy this for real — and who at your institution do you need on board first?

Model what you teach: disclose your AI useWhen your pipeline produces something students or colleagues see, say AI helped make it — briefly and plainly.

From draft to deployedNext steps: confirm your tools are institution-approved, pilot on de-identified or low-stakes data, keep the audit trail, and review with the right office (privacy / IT / accessibility) before it goes live.

Sources & Further Reading

Primary FERPA, 20 U.S.C. § 1232g; U.S. Dept. of Education, Student Privacy Policy Office (studentprivacy.ed.gov) — for any pipeline that touches student data.
Primary Section 508 of the Rehabilitation Act & WCAG 2.2 / ISO/IEC 40500:2025 (section508.gov) — for the accessibility of anything you deliver.
Primary Your institution’s AI / acceptable-use and data-governance policy — the authority on what you may deploy and on which tools.

A capstone pipeline is only deployable once its tools are approved and its data handling is cleared — verify both with your institution before going live.