Agentic Blueprints 36

A supervisor decides who works next — a researcher who searches or a coder who edits the repo — and assembles the result when the work is done.

Splits a hard question into sub-questions, researches them in parallel to keep the wall-clock low, then synthesizes one brief with citations.

Retrieval that doesn't trust its first search: it grades the results, rewrites the query when they're weak, and only answers once the context is good.

Turns a user story into browser tests: plans the scenarios, writes Playwright code, runs it against the real page, and self-heals until it passes for the right reason.

A failed data job triggers three probes in parallel, then a fixer and verifier loop on a patch until it passes a dry run — and open a PR.

Triages your inbox, then routes each thread to the right handler — draft a reply, propose meeting times, or file it — with your approval before anything sends.

A security alert gets enriched by three analysts in parallel, judged by a verdict agent with a skeptic on its shoulder, then auto-closed with rationale or escalated to a human.

Diff-aware multi-pass reviewer: finds bugs, then a skeptic agent verifies each before reporting.

Reads a GitHub issue and loops a coder against a verifier until it produces a minimal failing test that actually reproduces the bug.

A Sentry alert fans out to three investigators at once, then a commander writes the root-cause hypothesis and posts a status to Slack.

A planner breaks a goal into steps, an executor works them one at a time, and a replanner adjusts the plan as reality comes back — until it's done.

Answers questions about a codebase with citations to the exact file — and says so honestly when the answer isn't in the repo.

A new lead is enriched from the web and your CRM at once, scored against your ICP, and — if it fits — gets a tailored first email you approve before send.

A generator drafts, a critic tears it apart, the generator revises — repeat until the critic has nothing left to flag. Quality through self-review.

A scheduled sweep that reads competitor pricing, changelogs, and news in parallel, diffs against last week, and writes only what changed.

Turns a meeting transcript into a summary, then files the action items as tracked tickets and calendar follow-ups in parallel — and posts the recap.

Reads an incoming invoice, matches it against the purchase order and receipt in your database, and branches: clean ones queue for approval, mismatches get a written exception.

Extracts a contract's clauses, checks each against your playbook, and loops a redliner against a reviewer until it's clean — then a lawyer approves.

A scheduled sweep of CI history clusters intermittent failures, decides flake vs. real regression, and branches: quarantine or escalate.

Pulls Stripe charges and your warehouse's invoice records for the period in parallel, reconciles them line by line, and reports only the discrepancies — with likely causes.

Reads scattered user feedback in parallel, clusters it into themes with real quotes, and files the top requests as issues — with a digest.

Dry-runs a proposed SQL migration, checks it for locks and cost, and branches: safe ones get a green note, risky ones get a blocking report.

A reported email is dissected three ways at once — headers, links, attachments — then two agents debate benign vs. malicious before a human hits remediate.

English question to validated SQL to a plain-language answer. Planner, Analyst, Critic, Writer.