Research Desk turns sources into briefs. Outbid turns bids into a leaderboard. An AI coding platform turns ideas into running software — which makes it the strangest case in this class, because its output is other applications. Analyze it with the same card and it becomes a guided tour of everything Parts I–XI taught.

The user and the job

The user is someone with an idea for software and the authority to test it — a founder, a student, a team member — who wants to reduce the friction between the idea and a testable artifact. Not a demo video. Not a mockup. A running thing with files, a runtime, a preview, and a deploy path.

Apply the card's skepticism first. "Build an app" is not a wedge; it is a wish. The platform's real wedge is smaller: describe one outcome in plain language, get a running project with visible files you can inspect, adjust it conversationally, run it, and publish it. Each verb matters. The product keeps its promise only while the user can inspect, adjust, run, and deploy — remove any one and the path becomes a slot machine.

How Replit Agent works, from official docs

All capability claims below come from Replit's official documentation, verified September 2026. Interfaces and plan names change; the durable pattern — scope before changing, checkpoint everything, keep rollback one click away — is what you are studying.

Plan mode versus Build mode. Replit Agent offers two interaction modes that work best as a sequence. In Plan mode, the agent reads the project and produces guidance — ideas, task lists, success criteria, out-of-scope notes — but makes no file changes; it is read-only by design. In Build mode (the default), it writes code and modifies files directly. Approving a plan switches into Build automatically, and each completed task can create a checkpoint. The documented habits are explicit: be specific, plan the work, add context, review and test, use checkpoints — "agent can do a lot, but it works best when you lead" (Replit docs: Plan vs. Build Mode; Build with Agent).

Checkpoints and rollback. Checkpoints are automatic, complete snapshots of project state created at milestones — feature completion, major progress, stable tested states, before risky fixes. Each checkpoint preserves project files, installed packages and configuration, AI conversation context and agent memory, environment and publishing configuration, and database contents. Rollback restores the selected checkpoint with one click ("Rollback to here" from the Agent tab or history view), including conversation continuity so building resumes from the restored state; roll-forward through later checkpoints is also supported. Database restore is optional on rollback (selected explicitly; production databases restore separately), and every checkpoint also appears as a Git commit for branch-compatible version control (Replit docs: Checkpoints and Rollbacks).

Modes, spend, and deploy. Agent work runs under mode and spend controls: docs use tier names such as Free/Lite, Economy/Power, and Power/Max depending on page and date, with confirmations before paid actions and user-set usage limits so a loop cannot run up an unbounded bill. Tier names and credit amounts change frequently — use the live docs only for the mechanism, never quote a tier list as permanent. The same hosted path that builds the project also runs, previews, and deploys it.

Pricing and credit details (free allowances, per-tier credits) change frequently: check the live Replit pricing and Agent docs before quoting numbers. The mechanism — plan-then-build, checkpoint everything, one-click rollback, spend caps — is the stable lesson.

Mapping the stack

This case is a bridge back through the entire course. Map each layer:

chat interface → coding model → project files → runtime + preview
  → auth/integrations → test/reflection loop → deployment

The conversational interface (Part IX methods) takes intent. The coding model (Part IV) drafts implementation. Project files and checkpoints (Parts II, X) hold state and history. The runtime and preview (Parts III, VII) execute. Authentication and integrations (Parts V, XI) connect to users and services. The test and reflection loop plus checkpoints (Parts X, XI) verify. Deployment (Part VII) publishes. A failure at any layer presents as "the AI built the wrong thing," which is why the card's question 6 — what happens when wrong or unavailable — is the professional question: the answer here is inspect the files, read the plan, run the tests, roll back to the last good checkpoint, and fix forward or revert deliberately.

What the platform does not do for you

The docs are honest about this and so is this lesson. An AI coding platform accelerates implementation; it does not absorb responsibility. Five things remain yours no matter how smooth the demo:

1. Product definition. The wedge, the user, the non-goals — Classes 54–55 work. No agent can discover your user for you. 2. Security decisions. Secrets, permissions, ownership checks, exposed endpoints (recall Classes 47–48). Generated auth code still needs review. 3. Real testing. The agent can run checks, but acceptance against your user's actual job is your signature, not its summary. 4. User judgment. Whether the output is good, safe, and shippable is a human call. "The agent said it works" is not evidence. 5. Production responsibility. Logs, rollback readiness, data handling, support — Part XI in full. Deploying makes you the operator.

A practical rule: the platform handles work it can show you (files, diffs, previews, checkpoints); it merely hides work you cannot see unless you look (authz gaps, data exposure, silent failures). Your job is to look.

Hosted builder versus CLI/workspace: the control tradeoff

Earlier parts (III, X) set up this comparison. A hosted AI builder (Replit Agent) offers a managed all-in-one path: zero setup, shared runtime, built-in preview, deploy, and checkpoints. A CLI or workspace model (agent in your own repo and editor) offers ownership: your branches, your CI, your integrations, your cost controls, your exit path.

Neither wins generally. A weekend experiment or a first v1 with no infrastructure appetite favors hosted. A product with existing repos, compliance needs, custom infrastructure, or a team workflow favors the workspace. Choose per project, and write down what you traded.

Exercise: the recommendation memo

Analyze any AI app builder with the nine-question card, then write a half-page memo: which work it handles, which work it merely hides, which work remains yours — and your choice of surface for one named project with the control tradeoff stated.

Done means: a memo naming the project, the chosen surface, three handled/hidden/remaining items each, and one rollback plan ("if the next three tasks fail, we roll back to checkpoint X and fix forward by…"). Verify: someone else can implement your rollback plan from the memo alone. Common failure: choosing on vibes ("hosted feels easier") without naming what was traded. Convenience without a named price is how production surprises happen.

Check your understanding

• Why does Plan mode exist, and what goes wrong when builders skip it?
• What does a checkpoint capture that a Git commit alone does not?
• Name two responsibilities that stay with you after the agent writes the code.

Sources and unknowns

• Plan vs. Build behavior, checkpoint contents, rollback/roll-forward flow, Git integration: Replit official docs — Agent overview, Plan vs. Build Mode, Build with Agent, Checkpoints and Rollbacks (L2, verified September 2026 at https://docs.replit.com/features/agent/overview and https://docs.replit.com/features/version-control/checkpoints-and-rollbacks).
• Mode names (Free/Power/Max renames), spend limits, deploy path: Replit Agent and AI help docs (L2; plan names and credit amounts change — recheck live pricing before quoting).
• Unknowns (explicit): exact credit pricing and allowances at your read date; model identities behind each mode; production-database restore procedures beyond the documented separation — all undisclosed or time-sensitive here.

In the next lesson we study the trust layer every research product needs: source-linked answers, where citations are the product.