🤖

Agents & Automation

Break tasks → limit permissions → sandbox dry-run — agents act, you accept

Task

Break a repetitive job into agent steps → sandbox dry-run → accept output → name the reusable workflow

Time

~25 min

Suggested tool: Bolt.new

Self-rate (1–5)

How this scenario works

Agent failures usually mean too much permission or fuzzy success criteria — “clean up the data” becomes prod edits. This scene trains a verifiable automation loop: break repetitive work into ≤6 steps (input/action/output/check each) → cap folders and network → sandbox dry-run → human confirm before writes → save as skill/workflow → document failure fallback. Good for: CSV summaries, batch renames, weekly Markdown reports, in-repo scripting. Bad for: unaudited prod DB access, auto-send email/payments, pasting API keys into prompts. Example: job=summarize examples/feedback.csv into output/report.md daily; sandbox=read examples/ only, write output/ only, no network. Acceptable: step table with success criteria; after dry-run report.md has three agreed headings; you know which steps need human OK. Tool pick: Claude Code in the IDE; Devin for longer standalone jobs; Bolt.new for quick prototypes. Start small — verify on official sites.

Practice steps (with checks)

1Write the repetitive job + success criteria: input files, outputs, required structure. Done when you could manually QA the result without the agent.
2Inventory what exists: folders, commands, sample files; red-line bans (prod, network, secrets). Done when permission bounds fit on one screen.
3Run “Break into steps” for ≤6 agent steps (input/action/output/check/fallback each). Done when external-system steps are marked “needs human OK”.
4Human review: delete dangerous actions; batch writes to the last step if possible. Done when you approve before dry-run.
5Run “Sandbox dry-run” in [sandbox]; list paths before any write and wait for OK. Done when output exists and structure matches the deal.
6Accept against success criteria: headings, fields, no out-of-scope IO. Done when failure points to a step and rollback path.
7Run “Save workflow” with a name and how to trigger next time (command/skill/schedule). Done when a teammate could rerun from docs.
8Save step table + dry-run log; secrets always as placeholders. Done when similar automation starts in ≤15 minutes.

Practice Workbench

Repetitive jobSandbox boundaryStep listSuccess criteriaOutput sample

You are a senior automation engineer. Break this repetitive job into ≤6 agent steps:

Job: summarize examples/feedback.csv into output/report.md weekly
Sandbox: read examples/ only, write output/ only, no network

Rules: project files/commands only; no prod; no network; no real business data edits; each step has input/action/output/check; mark external steps “needs human OK”.

Markdown table: Step | Agent action | Permission | Success criteria | Failure fallback

Open Claude Code

Outcome checklist

Self rating (1–5)

More prompts (5)

Sandbox dry run

You assist Claude Code / Devin. Run a **minimal sandbox dry-run** from this step list:

Steps:
[steps]
Goal: [job]
Sandbox: [sandbox]

List paths before writes and wait; if files missing, list 3 samples to prepare; no out-of-scope paths.

Output: 1) Steps run 2) Output paths 3) Verification (structure/headings/errors) 4) Reusable workflow name

Permission boundary

Audit this agent plan for excessive permissions (review only — do not execute):

Job: [job]
Sandbox: [sandbox]
Steps:
[steps]

List: high-risk actions, steps to delete or mark human-OK, least-privilege rewrites (folders/network/commands).

Acceptance checklist

From the success criteria, make a human acceptance checklist for this agent output:

Goal: [job]
Expected structure:
[success_criteria]

Actual output summary (paste OK):
[output_sample]

6–8 checks: file exists, headings/fields, no scope creep, no invented data, rerunnable, rollback on failure.

Save workflow

Dry-run passed. Turn these steps into a reusable workflow doc (Markdown):

Job: [job]
Sandbox: [sandbox]
Steps:
[steps]

Include: workflow name, trigger (command/skill/manual), prerequisites, success criteria, failure fallback, variables to swap next time.
Do not recommend new tools or invent prices.

Paste safety check

Review what I’m about to paste to an agent for sensitive data (audit only):

[steps]

List secrets/tokens, internal hosts, prod paths, customer IDs. For each: delete/redact/OK. If none, say so — still remind me to human-check.

Recommended tools

Best fitFreemium

Claude Code

Anthropic's terminal-native AI coding assistant with deep codebase understanding, multi-file editing, test generation, and Git integration.

AlternativePaid

Devin

Cognition AI's fully autonomous AI software engineer that independently handles the full dev cycle from requirements to deployment.

Free-friendlyFreemium

Bolt.new

StackBlitz's browser-based AI full-stack app generator that creates runnable web apps from prompts with one-click deploy.

After you leave with templates,

ChooseNarrow tools with Choose LabBrowse guides & compares in Lab Related learning pathAgents & Automation Basics Try another scenario →

Continue your learning path

Practice a scenario Choose Lab

How this scenario works

Practice steps (with checks)

1Write the repetitive job + success criteria: input files, outputs, required structure. Done when you could manually QA the result without the agent.

2Inventory what exists: folders, commands, sample files; red-line bans (prod, network, secrets). Done when permission bounds fit on one screen.

3Run “Break into steps” for ≤6 agent steps (input/action/output/check/fallback each). Done when external-system steps are marked “needs human OK”.

4Human review: delete dangerous actions; batch writes to the last step if possible. Done when you approve before dry-run.

5Run “Sandbox dry-run” in [sandbox]; list paths before any write and wait for OK. Done when output exists and structure matches the deal.

6Accept against success criteria: headings, fields, no out-of-scope IO. Done when failure points to a step and rollback path.

7Run “Save workflow” with a name and how to trigger next time (command/skill/schedule). Done when a teammate could rerun from docs.

8Save step table + dry-run log; secrets always as placeholders. Done when similar automation starts in ≤15 minutes.

Practice Workbench

Repetitive jobSandbox boundaryStep listSuccess criteriaOutput sample

You are a senior automation engineer. Break this repetitive job into ≤6 agent steps: Job: summarize examples/feedback.csv into output/report.md weekly Sandbox: read examples/ only, write output/ only, no network Rules: project files/commands only; no prod; no network; no real business data edits; each step has input/action/output/check; mark external steps “needs human OK”. Markdown table: Step | Agent action | Permission | Success criteria | Failure fallback

Outcome checklist

Self rating (1–5)

You assist Claude Code / Devin. Run a **minimal sandbox dry-run** from this step list: Steps: [steps] Goal: [job] Sandbox: [sandbox] List paths before writes and wait; if files missing, list 3 samples to prepare; no out-of-scope paths. Output: 1) Steps run 2) Output paths 3) Verification (structure/headings/errors) 4) Reusable workflow name

Audit this agent plan for excessive permissions (review only — do not execute): Job: [job] Sandbox: [sandbox] Steps: [steps] List: high-risk actions, steps to delete or mark human-OK, least-privilege rewrites (folders/network/commands).

From the success criteria, make a human acceptance checklist for this agent output: Goal: [job] Expected structure: [success_criteria] Actual output summary (paste OK): [output_sample] 6–8 checks: file exists, headings/fields, no scope creep, no invented data, rerunnable, rollback on failure.

Dry-run passed. Turn these steps into a reusable workflow doc (Markdown): Job: [job] Sandbox: [sandbox] Steps: [steps] Include: workflow name, trigger (command/skill/manual), prerequisites, success criteria, failure fallback, variables to swap next time. Do not recommend new tools or invent prices.

Review what I’m about to paste to an agent for sensitive data (audit only): [steps] List secrets/tokens, internal hosts, prod paths, customer IDs. For each: delete/redact/OK. If none, say so — still remind me to human-check.