I summarized the most important lessons I learned after making over a dozen career moves, along with a simple framework to evaluate whether it is the time to stay in my current role or start looking for a new one.
https://t.co/NdzPiPHaxN
1/π§΅
@politicalmath Encoding the mechanics of a world model in a programming language is part of building software, but it is not the same as building software.
AI disproportionately favors the camp who can't tell the difference, erasing the correlation between good encoding and good engineering.
If I see "not x, not y, but z" I stop reading.
Guaranteed AI slop.
Not because it is artificial, not because it overuses an otherwise interesting writing style, but because it is usually trying to frame a bland argument as insight.
Everyone wants to contribute code.
Fewer want to integrate into team operations, task coordination, testing, migration, maintenance, operations, on call.
Writing the code is, by large, the fun part of the work.
Volunteering only to do the fun part is ... a choice.
@robertgraham The reporting is correct: people in the Philippines are helping guide the cars.
"But is not that much" is a weird counter-argument. The range of assist and conditions is behind a black box and can change at a moment's notice.
The point is: there is a human-driver involved.
@jamesacowling We are quickly careening off towards gigantic code-bases that no one can hold in their brains anymore.
I started seeing regular PRs with 10k line changes with 80% boilerplate code of redundant input sanitization and exception catches across 4-5 levels of nested calls.
@erenbali Can we add some basic controls to prompts?
Nothing fancy, just some syntax for looping through instructions.
Oh, and maybe a way to give a name to some of prompt sequences so it is easier to reuse them?
@thebiryanidev Acting like an owner means thinking through consequences to the business, not working longer hours.
I have seen people working long hours and executing bad direction as a matter of course.
Being average on hours is healthy. Being average on ownership may become an identity.
@mmatthias@samswoora Isn't that like saying anyone can be a doctor without experience or skills but they need to know what to ask patients?
The "but" part is the experience and skills.
Growth papers over a lot of bad business cases.
Lack of growth highlights all that is wrong with broken business cases.
That also works with society.
That also works with life.
The universe demands growth.
I wonder how much of the "programming is dead" discourse is the result of people who never worked on production system being able to create prototypes.
I can ask ChatGPT for good medical opinions all day and still would not think I can handle a general population of patients.
Using "coding" and "engineering" interchangeably does not help the discourse.
An AI writing code has no concern for engineering across the entire stack from requirements management to monitoring and supporting production code.
Aligning all those processes remains unsolved.
Compound Engineering is what happens when agents write 100% of the code.
At @every, engineers donβt type code anymore. They orchestrate agents.
The shift:
- Coding is no longer the bottleneck
- Planning, review, and learning loops matter more than syntax
- Each feature makes the next one easier to build
The 4-step Compound Engineering loop:
1.Plan β Agents research the codebase + best practices and produce detailed plans
https://t.co/fMdbK6qY7v β Agents write code, tests, and iterate using real app simulations
3.Assess β Humans + AI review from multiple angles (security, performance, overbuild)
4.Compound β Lessons learned are stored so future agents never repeat mistakes
Complexity still grows, but so does the AIβs understanding of the system.
Result:
- One developer can now do the work of ~5
- Products run by single engineers serve thousands of users
- New hires instantly inherit years of institutional knowledge
Engineering is no longer about writing code.
Itβs about designing learning systems that compound.
β Thanks @danshipper and @kieranklaassen for sharing your approach. So much to learn and takeaway from this compound engineering approach.
Compound Engineering is what happens when agents write 100% of the code.
At @every, engineers donβt type code anymore. They orchestrate agents.
The shift:
- Coding is no longer the bottleneck
- Planning, review, and learning loops matter more than syntax
- Each feature makes the next one easier to build
The 4-step Compound Engineering loop:
1.Plan β Agents research the codebase + best practices and produce detailed plans
https://t.co/fMdbK6qY7v β Agents write code, tests, and iterate using real app simulations
3.Assess β Humans + AI review from multiple angles (security, performance, overbuild)
4.Compound β Lessons learned are stored so future agents never repeat mistakes
Complexity still grows, but so does the AIβs understanding of the system.
Result:
- One developer can now do the work of ~5
- Products run by single engineers serve thousands of users
- New hires instantly inherit years of institutional knowledge
Engineering is no longer about writing code.
Itβs about designing learning systems that compound.
β Thanks @danshipper and @kieranklaassen for sharing your approach. So much to learn and takeaway from this compound engineering approach.
@plk669888@burkov Chores.
- "analyze code changes, create a commit message following project guidelines at this URL, create a pull request following project template"
- "look at this PR from colleague, create feedback for disciplines X, Y, Z. Inspect adherence to project guidelines."
...