Course Content
Agentic AI Patterns
9 sections · 50 lessons
Elaborate on a few real-world applications of Agentic AI.
What you need to know
The test for a good agent use case
Ask three questions:
- Does it need several steps whose order depends on what is found? If the order is fixed, use a workflow.
- Does it need tools? Live data, private systems, or actions.
- Can success be checked? Tests pass, a record reconciles, a human approves, a ticket stays closed.
The third question matters most. An agent improves when it can see that it failed: a test turns red, an API returns an error, a total does not match.
Applications and their success signals
| Application | What the agent does | How success is checked |
|---|---|---|
| Coding agents | Read repo, edit files, run tests, iterate | Test suite passes, PR review |
| Customer support | Look up order, apply policy, refund or escalate | Ticket not reopened in 7 days |
| Deep research | Search many sources, read, synthesise | Citations verify; expert review |
| Insurance claims | Gather documents, check policy, draft decision | Adjuster agrees with draft |
| DevOps incident triage | Pull logs, metrics, recent deploys; suggest cause | Matches post-incident root cause |
| Procurement | Collect vendor quotes, normalise, compare | Buyer accepts the comparison |
| Travel booking | Search flights and hotels, hold, book | Booking confirmed within constraints |
| Computer-use agents | Operate a GUI from screenshots | Form submitted, record visible |
Computer-use agents see the screen as images and send mouse and keyboard actions. They reach old systems with no API, but they are slower and less reliable than API tools, so use them only when no API exists.
Where agents disappoint
- Open-ended strategy or creative work with no checkable outcome.
- High-volume tasks with a fixed path, where a workflow is 5 to 10 times cheaper.
- Real-time control loops that need millisecond responses.
A real-life example
A mid-size logistics company runs 60 microservices. At 2 a.m. an alert fires: payment API p95 latency jumped from 300 ms to 4 s.
The DevOps triage agent does what the on-call engineer used to do in the first 20 minutes. It calls get_recent_deploys (a config change went out 12 minutes earlier), query_metrics (database connection pool at 100%), and search_logs ("too many connections" errors starting 11 minutes ago). It posts a summary in the incident channel: likely cause, evidence links, and a suggested rollback command.
It does not run the rollback. A human does that after reading the evidence. The team measures the agent by whether its suggested cause matched the root cause in the post-incident review; it matched in about 7 of 10 incidents in their first quarter, and cut time-to-first-hypothesis from about 20 minutes to 3.
Follow-up questions to expect
- "Why are coding agents ahead of other domains?" — Code has cheap, fast, automatic verification: compilers, linters and tests. The agent gets clear error messages and can retry.
- "Would you build an agent for sending marketing emails?" — Mostly no. Drafting is one model call, and sending should be a human-approved step. The step order is fixed, so a workflow fits.
- "What is a computer-use agent good for?" — Legacy systems with no API, such as an old insurer portal. For anything with an API, a typed tool is faster, cheaper and more reliable.