Skip to content

Learn · Track

AI agents by job

One job per guide: what goes in and comes out, a diagram of the design, which tools the agent may read and which need a person's approval, how to test it with a golden set, what drives its cost, and the sample or recipe to start from. Each guide says plainly which parts the sample covers and which you build yourself.

Guides in this track

Intermediate12 guides, in reading order

  1. Support ticket triage

    An agent that reads new tickets, sets urgency and topic, drafts a reply and routes it, with a person approving anything before it reaches a customer.

    9 min read

  2. Incident triage

    Gather evidence from logs, metrics and recent changes over MCP, summarise impact, open a ticket and page a person, with read-only tools and approved fixes.

    10 min read

  3. Data reporting agent

    Build a scheduled report: query a database read-only, chart it with Vega-Lite, let a model write the summary, and check its numbers before it is sent.

    10 min read

  4. Internal docs Q&A

    Answer staff questions from your own documents with a self-hosted knowledge base, hybrid search and citations, and keep restricted files out of answers.

    9 min read

  5. Security questionnaires

    Draft answers to vendor security questionnaires from your own policies with citations, flag what they do not cover, and have an owner approve every answer.

    9 min read

  6. Code review agent

    Run an agent on each pull request, let it read the diff with read-only MCP tools, post one review comment, and leave approving and merging to people.

    11 min read

  7. Refund approval

    An agent proposes a refund, a person approves the exact amount, the payment call runs once with an idempotency key, and the decision is signed and logged.

    10 min read

  8. Research with citations

    A supervisor and worker agents research a question over your sources and return claims with checkable citations, plus a test that catches made-up sources.

    8 min read

  9. Email to structured data

    Have a mail relay post emails to a webhook, extract fields into JSON with a schema, check them in code, write clean rows, route doubtful ones to a person.

    9 min read

  10. Scheduled monitoring agent

    Run a check on a cron schedule, compare the result with the last run, and alert only on real changes. How to build the comparison in L2M, and cap its cost.

    9 min read

  11. Slack assistant

    A Slack-triggered agent that keeps memory per thread, uses a short list of MCP tools, replies in the thread and asks a person before any write action.

    10 min read

  12. CRM enrichment agent

    Enrich new CRM records from a company's public website, propose field updates as checked JSON, and write them only after a person approves, under GDPR.

    10 min read

Before and after

Prerequisites: what an AI agent is and human in the loop. The pattern library shows the building blocks these designs combine.

After this track: import a working starting point from the recipes, and read evals for AI agents before you put one in front of customers.

This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings