firaz.
← writing
agentsJun 4, 2026·2 min read

the best way to actually learn ai agents in 2026 (from someone shipping them daily)

the honest path isn't a course or a framework tier-list. it's building one ugly agent that does one real thing, in production, this week. here's the route i'd take if i started over.

"what are the best resources to learn ai agents in 2026?" gets asked constantly, and almost every answer is a list of courses, frameworks, and youtube channels. that's the slow road. i ship agents every day, and if i were starting over, i'd skip almost all of it.

here's the route that actually works.

build one ugly agent that does one real thing — this week

not a tutorial agent. a real one, for a real annoyance in your own life or work. "summarize my morning emails." "draft replies to inbound dms." "post my notes to a doc." one job, in production, used by you, by friday.

you don't learn agents by studying agents. you learn them by being on the hook for one that has to work tomorrow.

the courses teach you the vocabulary. the ugly agent teaches you the thing courses can't: what breaks, why, and how it feels when something you built keeps running without you.

then learn these four things, in this order

  1. the loop. an agent is just: get context → call the model → take an action → check the result → repeat. understand that loop cold. every framework is a wrapper around it.
  2. tools/function calling. the model deciding which action to take is where agents stop being chatbots. this is the real unlock.
  3. memory and context. what the agent knows, and how it stops forgetting. this is where most agents get good or stay useless.
  4. making it maintainable. logging, versioning, constraints. the boring part that separates a demo from a system. learn it early, not after it breaks.

skip these (for now)

the framework tier-lists. the "agentic AGI" hype. the 8-hour courses. you'll absorb what you need from them after you've shipped one — and then they'll actually make sense instead of being abstract.

the meta-move

i ship one open-source tool a day. that constraint taught me more than any course, because it forces the only thing that matters: you have to finish and ship something real, repeatedly. pick a version of that for yourself — one agent a week, in public. the public part isn't vanity. it's the deadline that makes you actually do it.

the people who "know ai agents" in 2026 aren't the ones who watched the most. they're the ones who shipped the most. start ugly. start this week.

written by firaz fhansurie