Use Cases
Every scenario on this page is buildable with Agent Ops as it is today — the persistent sandbox, git tools, schedules, webhook automations, actions, skills, approvals, and external-event waits. Each one lists the setup it needs and an example task prompt you can adapt.
A useful mental model for reading them:
Automation decides when work starts · Template defines what it is · Skills define how it's done · Actions are the systems it touches · Approvals are where a human signs · Memory makes run n+1 better than run n.
Engineering & DevOps
Automated pull request review
Setup: GitHub action · pull request opened automation · GitHub PR review template · a code-review skill with your team's standards · GITHUB_TOKEN env var.
When a PR opens, a task is created automatically. The agent clones the branch, reads the diff in context (not just the patch), applies your review skill, and posts comments through the GitHub action. Merging stays human.
Review PR #{{number}} in {{repo}}. Check correctness, security, and
adherence to our review skill. Post inline comments for specific issues
and a summary comment with a merge recommendation.Issue → first-draft implementation
Setup: GitHub action · issue created automation · GITHUB_TOKEN.
New issues become tasks; the agent reproduces the problem where possible, drafts the fix on a branch, and pushes it for review — the push itself pauses for approval, so nothing lands unseen.
Nightly test-coverage patrol
Setup: cron schedule (nightly) · GitHub action · GITHUB_TOKEN.
Clone {{repo}}, find the three modules with the weakest test coverage,
write meaningful tests for them, run the suite until green, and push a
branch named tests/nightly-{{date}}. Summarize what the new tests cover.The morning email tells the team what arrived overnight; the push approval is the human gate.
Incident triage and postmortem drafting
Setup: webhook automation wired to your alerting · Slack action · incident-triage and postmortem templates.
An alert POSTs to the automation URL with the incident payload. The agent gathers context, posts a first-status message to the incident channel, and — as a follow-up task — turns the resolved incident's notes into a structured postmortem draft.
Dependency update patrol
Setup: weekly cron · GitHub action · web search (built in).
The agent lists outdated dependencies, reads their changelogs on the web, separates safe patch bumps from breaking upgrades, applies the safe set, runs the tests, and opens a branch — with its reasoning in the task log. Memory keeps track of packages you've deliberately pinned.
Long-running job with CI feedback
Setup: external events (built in) · your CI's webhook pointing at the task's callback URL.
Push the migration branch, then wait for our CI to report the result
to your callback URL. If the pipeline fails, read the log excerpt from
the event payload, fix the issue, and repeat — up to three attempts.This is the external-event primitive doing real work: the run suspends (up to seven days) between attempts and consumes nothing while CI runs.
Codebase onboarding guide
Setup: GitHub action or just git clone · nothing else.
Clone {{repo}} and write ONBOARDING.md for a new backend developer:
architecture map, how a request flows from route to database, where
configuration lives, how to run it locally, and the five files most
worth reading first.Multi-repository migration with subtasks
Setup: Allow child tasks on · GitHub action · GITHUB_TOKEN.
A parent task plans the migration and proposes one subtask per repository (each creation approval-gated). Each subtask applies the change and pushes a branch; the parent waits for all of them (wait for task) and assembles a status report of what merged cleanly and what needs a human.
Finance
Monthly management report
Setup: monthly cron · SQL action to your reporting database · a report-format skill · project email.
Build the monthly management report for {{month}}: pull revenue, costs,
and headcount from the finance DB, compute deltas vs. last month and
vs. budget (budget.xlsx in the workspace), produce the standard workbook
per the report-format skill with charts, and email it to finance@acme.com.The persistent workspace keeps every prior month's workbook, so trends and formats carry forward. This is the workflow class where teams report cutting a 30-hour monthly close-support grind to a fraction.
Variance bridge with human judgment
Setup: SQL action or uploaded exports · a variance-policy skill.
The agent reconciles budget to actuals line by line and labels every variance — but it doesn't decide. Unsupported variances become a question:
Marketing spend is +18% with no plan change on file. Adjust the
forecast, or hold and escalate to the budget owner?The run parks as waiting for input, the owner gets the email, and the report finishes with the human's call recorded in the run history.
Invoice intake and reconciliation
Setup: IMAP/M365 email action · SQL action · a reconciliation skill.
On a daily schedule the agent reads the invoices mailbox, parses PDF attachments (pdfplumber is preinstalled), matches each invoice against open purchase orders in the database, and files a mismatch report. Nothing is auto-posted to accounting — the deliverable is the reconciled list plus exceptions for a human.
Subscription & payments health report
Setup: weekly cron · Stripe (or iyzico/PayU) action · Slack action.
Failed payments, involuntary churn candidates, plan-change trends — summarized weekly to the finance channel, with the raw table attached as an Excel file.
Board deck data pack
Setup: SQL action · a slide-format skill.
Prepare the Q2 board data pack: KPI summary slide, revenue bridge,
cohort retention chart, and cash position, as board-Q2.pptx following
the slide-format skill. Flag any metric that moved >10% QoQ with a
one-line explanation.Legal & Compliance
Contract review with risk labels
Setup: a contract-review skill encoding your standard positions.
Upload a contract (docx or PDF) and run:
Review acme-msa.docx against our contract-review skill. Produce the
clause table with GREEN/YELLOW/RED labels and a recommendation memo.
Escalate anything RED as a question rather than deciding.NDA fast lane
Setup: contract-review skill · webhook automation from your intake form (optional).
Incoming NDAs are diffed against your standard template. Matches get an "approved as standard" memo in minutes; deviations park as a question for the lawyer with the specific clauses quoted. The lawyer reviews exceptions, not the pile.
Bulk clause extraction across a contract archive
Setup: upload the archive (or point at a synced folder) · nothing else.
Across every contract in /workspace/contracts: extract party, effective
date, term, renewal mechanics, and liability cap into one Excel register.
Flag contracts with auto-renewal windows closing in the next 90 days.Classic batch-document processing — hundreds of files, one pass, one register.
Marketing content legal pre-review
Setup: a legal-review skill with your claims policy · webhook from the marketing team's intake form.
Submitted copy is checked against the claims policy; the agent suggests compliant rewording for common issues and escalates genuinely ambiguous claims to counsel. Legal reviews the hard 10%, not all 100%.
Compliance calendar
Setup: monthly cron · project email.
Statutory deadlines (filings, renewals, data-protection reviews) live in a workbook in the workspace; the agent checks it monthly, emails owners of anything due within 60 days, and records confirmations it receives back via task comments.
HR & Recruiting
CV screening with a human short-list
Setup: a role-requirements skill per open position · applications uploaded or webhook-fed from your ATS.
The agent scores each CV against the requirements skill and produces a ranked table with per-candidate rationale. It never rejects anyone — the deliverable is a short-list proposal; the recruiter's decision closes the task.
Interview brief pack
Setup: web search (built in) · calendar/ATS data via webhook or upload.
For tomorrow's interviews (list attached): for each candidate, produce a
one-page brief — CV summary, public profile highlights, match against the
role skill, and five tailored interview questions. One PDF per interviewer.Offer letter production
Setup: an offer-template skill · project email · approvals on.
Approved candidate details go in; a formatted offer letter (docx) comes out. Sending the email is the approval gate — HR signs the send, not the typing.
Onboarding task fan-out
Setup: webhook from your HRIS ("new hire") · Jira/Asana action.
A new-hire event creates an onboarding task: the agent opens the IT, payroll, and buddy-assignment tickets in your tracker, generates the first-week schedule document, and reports what's pending.
Sales
Pre-meeting account brief
Setup: CRM action (HubSpot/Salesforce/Pipedrive) · web search · Slack action.
Prepare a brief for tomorrow's Acme renewal call: last five CRM
touchpoints, open tickets, current plan and usage, recent company news,
and our win themes vs. their shortlisted competitor. Post to #sales-briefs.Run it manually before big calls or on a daily 07:00 cron against the day's calendar export.
Post-call follow-through
Setup: CRM action · Jira/Asana action · approvals on for CRM writes.
Upload the call notes; the agent drafts the follow-up email, proposes the CRM field updates (approval-gated), and opens the committed action items in your tracker.
Pipeline hygiene patrol
Setup: weekly cron · CRM action · Slack action.
The agent flags stale opportunities, missing next steps, and slipped close dates — and proposes cleanups rather than performing them. Destructive CRM operations stay behind approval, which is exactly what prevents the classic "clean up old opportunities" agent accident.
Competitor watch
Setup: recurring competitor-watch template · cron · web search · memory.
Because memory keeps what was seen before, the weekly report contains only changes: new pricing, new features, new positioning — not a re-description of the competitor's website.
RFP first draft
Setup: a company-facts skill (approved boilerplate, security answers, references) · the RFP uploaded.
The agent parses the RFP question by question, answers what the skill covers, and produces a draft with every unanswered or low-confidence item explicitly listed for the humans — a triaged document instead of a blank one.
Marketing
Ad variation factory
Setup: performance export (CSV) uploaded or webhook-fed · a brand-voice skill with hard character limits.
From ad-performance.csv: identify creatives below the account-average
CTR, and for each generate 20 new headline/description variants within
platform character limits, in our brand voice. Deliver as one CSV per
campaign, ready for bulk upload.The scenario the "one-person growth-marketing team" story is made of: hours of copy-paste becomes minutes of review.
Weekly campaign performance report
Setup: weekly cron · channel exports (upload or API via HTTP) · Slack action.
Cross-channel spend, ROAS, and CAC in one workbook with charts, plus a three-bullet narrative to the marketing channel. Memory keeps the agreed formatting and last week's numbers for comparison.
Content calendar execution
Setup: weekly cron · content calendar workbook in the workspace · brand-voice skill · approvals on sends.
Each week the agent drafts the scheduled pieces in your voice and parks them for editorial approval — the editor approves or annotates, and approved pieces move on (email via SendGrid/Mailchimp action where that's the channel).
Campaign brief generator
Setup: webhook from a simple intake form · a brief-format skill.
Form fields in (campaign name, audience, key message, channels) → a standard creative-brief document out, filed in the project and linked back to the requester. The classic "lightweight internal tool" — built as configuration, not as software.
Data & Analytics
Self-service data questions
Setup: SQL action (read-only credentials recommended) — that's it.
Why did weekly active users dip in the last two weeks? Break it down by
platform and acquisition channel, chart the trend, and tell me whether
it's seasonal by comparing with the same weeks last year.Business users get an analyst-grade answer — query, chart, and narrative — without an engineering ticket. Read-only mode makes this project safe to share widely.
Metric anomaly sentinel
Setup: daily cron · SQL action · Slack action · memory.
The agent pulls the KPI set daily, compares against remembered baselines, and stays silent when everything is normal — posting to the channel only when something deviates, together with its first-pass investigation of why.
Data quality audit
Setup: data-quality-audit template · weekly cron · SQL action.
Null rates, duplicates, referential breaks, schema drift — trended over time in the workspace, so the report shows whether quality is improving or eroding.
Ad-hoc file crunching
Setup: none — upload and ask.
Pivot a messy export, merge twelve regional workbooks into one, convert a folder of PDFs into a structured table. pandas + openpyxl + a persistent workspace cover the everyday "can someone process this file" traffic.
Operations & Administration
Weekly team status report
Setup: weekly cron · Jira + GitHub + Slack actions · a status-format skill · project email.
The agent compiles the week from your systems — shipped, slipped, blocked — into the exact format your leadership reads, and emails it Friday afternoon. The first two weeks you correct it; memory does the rest.
Meeting notes → tracked actions
Setup: Jira/Asana action · a follow-up cron.
Upload meeting notes; the agent extracts decisions and action items, opens tracker tickets with owners and due dates (creation approval-gated if you prefer), and a weekly cron nudges owners of anything still open.
Vendor evaluation dossier
Setup: web search · an evaluation-criteria skill.
Evaluate these four vendors for our helpdesk migration against the
criteria skill: feature fit, pricing (as published), security posture,
references. Produce a comparison matrix and a two-paragraph
recommendation with the key risk for each finalist.Document intake triage
Setup: scanned documents uploaded (or webhook-fed) · a routing skill.
Each incoming PDF is classified (invoice / contract / official notice / other), renamed per your convention, filed in the right workspace folder, and — where the routing skill says so — forwarded to the right owner by email.
Waiting on the outside world
Setup: external events (built in).
Any process that hands work to a third party — a vendor approval, an official application, a partner's batch job — can be modeled as: do the preparation, send it out with the callback URL, suspend up to seven days, resume when the answer arrives (or the timeout forces a follow-up). The task's run history becomes the audit trail of the whole exchange.
Customer Support
Ticket enrichment and triage
Setup: webhook from your helpdesk · CRM action · memory.
New tickets arrive with context attached: the customer's plan and history from the CRM, similar past incidents from project memory, and a suggested first response — the human agent starts from a dossier, not a blank reply box.
Draft-then-approve responses
Setup: approvals on · email action.
For sensitive queues, the agent drafts every response and a human approves each send. It's the standard bridge to automation: measure how often drafts go out unedited before deciding what to loosen.
Knowledge-base gardening
Setup: weekly cron · helpdesk export or action.
Resolved tickets are mined for missing articles; existing articles that contradict recent resolutions get flagged as stale. The deliverable is a weekly KB changelist for the support lead to accept.
Patterns to steal
Four shapes recur across everything above — recognize them and you can design your own scenarios:
- Batch document processing — hundreds of files in, one register/report out. Needs only the sandbox.
- Recurring report with memory — cron + data action + format skill; quality compounds weekly.
- Webhook → template → approval — external event in, standardized task out, human signature exactly where policy demands it.
- Prepare → wait → continue — external-event waits for anything that takes days: CI, vendors, officials, other teams.
Start with one project, one task, everything gated — then remove the gates the run history proves you don't need.