Muhammad Haroon

AI Engineering Educator. Hands on cohorts where you build real agents, apps, pipelines. For engineers, switchers, tech leads. DM cohort

[Muhammad Haroon @Muhammadh811]

I help engineers, switchers and tech leads learn AI by building real systems.

Not slides. Agents, apps, pipelines.

Hands on cohorts. If you want in, DM cohort

[Muhammad Haroon @Muhammadh811]

@Nik_Coding10 same. once the need is clear the slow path usually finishes sooner

[Muhammad Haroon @Muhammadh811]

students ask which model is best

i ask which one their company already pays for

that usually settles it

[Muhammad Haroon @Muhammadh811]

every AI tutorial ends the moment the code runs

every real job starts there

[Muhammad Haroon @Muhammadh811]

the most useful AI skill nobody posts about is reading the logs

the model did something dumb
the log says why

[Muhammad Haroon @Muhammadh811]

every agent demo says human in the loop

the human is me clicking approve without reading

[Muhammad Haroon @Muhammadh811]

@yurisa_xi model switching in one panel is handy. i'd pin the model per task though, otherwise you can't tell why a result changed between runs

[Muhammad Haroon @Muhammadh811]

@aiDotEngineer @AlixPartnersLLP counting savings twice is the kind of mistake a person spots in 5 seconds and an agent never questions. it needs a check step before the number goes to anyone

[Muhammad Haroon @Muhammadh811]

@EveryDevAi @earthtojake i started in mechanical engineering, so the DFM checks are the bit i'd test first. a part that looks right but can't be made is the classic mistake

[Muhammad Haroon @Muhammadh811]

@RoundtableSpace 30 agents at once is easy. reviewing 30 diffs after is the real job. the plan step before it runs is the part worth keeping

[Muhammad Haroon @Muhammadh811]

@p1xel88ted a chief bot that only routes the work is a good pattern. what happens when one of the worker bots gets a task wrong, does the chief catch it?

[Muhammad Haroon @Muhammadh811]

@BIGMayrr most people dump the whole repo into context when 20 lines about what not to touch would do more. context quality beats context size

[Muhammad Haroon @Muhammadh811]

@CuriousDevX the same agent that helps a pentester helps an attacker. the boring control left is logs of what it ran and who started it

[Muhammad Haroon @Muhammadh811]

@MaxDevsLabs @Teknium @bcherny @thsottiaux @_mohansolo keeping the next action in git is a good call. handoffs between agents usually break because nobody wrote down what was already tried and failed

[Muhammad Haroon @Muhammadh811]

@naveen_saradhi_ a short CLAUDE.md with the real decisions in it fixes maybe half of this. curious what claude-mem catches that a hand written file would miss

[Muhammad Haroon @Muhammadh811]

@EricWangclaude same model reviewing its own work shares the same blind spots, so extra rounds add nothing. makes sense that a different model in the loop moves the number

[Muhammad Haroon @Muhammadh811]

@melvindvivas reading the readme was never the part that mattered. skimming what the install script touches on your machine is. i'd still read that bit before the agent runs it

[Muhammad Haroon @Muhammadh811]

the agent passed every test
i let the agent write the tests

great team

[Muhammad Haroon @Muhammadh811]

power automate is the most underrated AI tool on here

in a big company it already has approvals and IT sign off

your langgraph demo has a readme

[Muhammad Haroon @Muhammadh811]

8 people on X i'd follow before buying any AI course:
1. @karpathy
2. @simonw
3. @swyx
4. @eugeneyan
5. @HamelHusain
6. @jxnlco
7. @hwchase17
8. @levelsio (he'll tell you to skip the course anyway)

[Muhammad Haroon @Muhammadh811]

@icyklop1 65 agents and ~9,000 tool calls in 11 hours is wild. how much of the editor's time went into checking the output after?