Param Harrison

Founder @ | Agentic AI & Software Delivery Transformation | Teaching engineers to build production AI systems ↓

[Param Harrison @learnwithparam]

@ahmedgagan11 Put it in trustMRR for reach Ahmed 😁

Congrats on the launch, looks amazing, hope it works amazing too

[Param Harrison @learnwithparam]

Want to build better systems around your coding agents?

Join my free lightning session: 👇

🚀 𝗕𝘂𝗶𝗹𝗱 𝗮 𝗠𝗶𝘀𝘀𝗶𝗼𝗻 𝗖𝗼𝗻𝘁𝗿𝗼𝗹 𝗳𝗼𝗿 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀

https://t.co/WesIeNhwzK

[Param Harrison @learnwithparam]

Most coding agents guess how software works.

REA gives them tools to investigate it.

4 lessons for engineers:
→ Investigate before implementing
→ Give agents tools, not just prompts
→ Demand evidence
→ Separate discovery from implementation

https://t.co/q7e9E6Ll8o

[Param Harrison @learnwithparam]

@spacedevin https://t.co/7pmf5H33W9 Learn to speak english without any hesitation using the AI tutor

[Param Harrison @learnwithparam]

@hanghuang_ No doubt, that seems a great offer, but at what cost. 20vCPU and 40GB? How are you funded and do you keep the data for analysis on the invocations?

Any terms to worry on before running our tests :)

BTW, I teach at Maven around software factory which has sandbox execution.

[Param Harrison @learnwithparam]

Top 1% Engineers, my free newsletter: https://t.co/CoF7Y99Hsx

[Param Harrison @learnwithparam]

The agent passed every test. Because it wrote them.

The fix: static gates before review, holdout tests outside its worktree, and a loop that sends failures back with the reason.

[Param Harrison @learnwithparam]

Start this week: move 5 behaviour tests out of the agent's reach and watch its next PR.

[Param Harrison @learnwithparam]

Build your own Software Factory
https://t.co/1Ex0xMdMWx

[Param Harrison @learnwithparam]

Your safety hook crashed. The agent ran the command anyway.

In Claude Code, exit 1 and timeouts don't block. Deny with exit 2, add onFailure: "block" (2.1.295), then break the hook on purpose.

[Param Harrison @learnwithparam]

@navuud @swombat Dude, they are asking proof of the software factory and PRs which you merged.

Not some random ads manager pages when the post is about software development acceleration.

I myself teach building software factory at maven. I have no problem trusting the post until random video

[Param Harrison @learnwithparam]

@bzagrodzki @openseo_so Is it really 0 per month or you have hidden cost to API like dataforseo?

Still, it is cheap and best to have the open source But that $0 per months seems misleading

[Param Harrison @learnwithparam]

My next live session on Maven
https://t.co/nHaxxW0syR

[Param Harrison @learnwithparam]

https://t.co/kqLTiwYwXf

[Param Harrison @learnwithparam]

The agent gives you the code. The workflow decides if it ships.

Issue → task with a done check → one isolated agent → checks before people → PR to a named reviewer.

Draw it for your last 5 PRs.

[Param Harrison @learnwithparam]

https://t.co/FRMIfzDp0q

[Param Harrison @learnwithparam]

When agents open the pull requests, your best reviewer becomes the queue. Count reviews per person for 30 days, name a second owner per risky path, let checks go first, and cap open agent changes at what your team can review.

[Param Harrison @learnwithparam]

I build my brain 😉 or as they call it second brain.
I would say, everyone should build one to keep track of how they consume or produce content. The brain dump outside of your brain will help to learn some of the concepts of AI engineering in depth while building.

[Param Harrison @learnwithparam]

A one-line ticket goes in. Three guesses come out. Before any agent starts, add five lines: the outcome, three acceptance checks, what is out of scope, the paths it may touch, and the owner who approves the release.

[Param Harrison @learnwithparam]

AI coding agents don't need to get better at writing code.

They need to get better at proving they're right.

Separate the Doer from the Tester.

Doer → Verify → Test → Fix → Verified PR

Build a Claude Code Verification Harness
https://t.co/9vlV6As2UU