Adam West

i build useful software

[Adam West @adamscodex]

There Is No Anti Memetics Division

Best book I've read since The Dark Forest

[Andrew Côté @Andercot]

This is a computer

[Adam West @adamscodex]

@antirez https://t.co/pwxPPWJLjK

[Adam West @adamscodex]

How to keep codex and claude code running all night to make your dreams come true:

codex> what’s the next step to take this idea to production ready without me I’m busy?

*tab*

codex> do it and then commit

*tab*

codex> what’s the next step to take this idea to production ready without me I’m busy?

*tab*

codex> do it and then commit

*tab*

Now do that 20x times.
Then “summarize the last 20 commits”
Prune the bad ideas, keep the good ones.

Gotta love codex infinite pro plan limits😂
get it while you can.

[Adam West @adamscodex]

How to keep codex and claude code running all night to make your dreams come true:

codex> what’s the next step to take this idea to production ready without me I’m busy?

*tab*

codex> do it and then commit

*tab*

codex> what’s the next step to take this idea to production ready without me I’m busy?

*tab*

codex> do it and then commit

*tab*

Now do that 20x times.
Then “summarize the last 20 commits”
Prune the bad ideas, keep the good ones.

Gotta love codex infinite pro plan limits😂
get it while you can.

[Adam West @adamscodex]

@Andercot seems sketchy

[Adam West @adamscodex]

thinking for yourself is now rarer than ever

[Vlad Khononov @vladikk]

Remember the microservices hype 12 years ago?

All the cool kids couldn't stop talking about how microservices were the Holy Grail of software engineering. Ironically, most of them couldn't tell you what a microservice actually was. Like a bunch of blind men describing an elephant.

How that turned out? First, the comeback of the modular monolith, as half the crowd quietly admitted they never needed the network calls in the first place. Second, the microservices heroes mostly ended up with distributed monoliths. And to cope with the insane web of dependencies running across their poorly drawn service boundaries, they came up with a solution: the monorepo.

Nowadays, the cool kids discovered the new Holy Grail: the AI (s)loop. New protagonist, same plot: a crowd chasing something most of them can't quite define. 10 years ago it was about thousands of services, today it's thousands of deploys per day. Let's see what we'll reach for to cope with the hangover this time

[Adam West @adamscodex]

The biggest opportunities are in domains where mistakes are expensive: rockets, robotics, infrastructure, security, healthcare, science, and finance.

[Adam West @adamscodex]

Robotics needs CI/CD for the physical world

[Adam West @adamscodex]

The most underrated agent-first product feature is not chat.

It’s dry-run.

Before an agent changes anything, it should be able to ask:

“What would happen if I did this?”

Then get a structured answer.

That is how autonomy becomes safe enough to use.

[Adam West @adamscodex]

@Andercot We’re going back to the future…again

[Adam West @adamscodex]

Robobun is interesting because it doesn’t just generate code.

It closes the loop.

Reproduce the issue.
Prove the bug.
Make the change.
Prove the fix.
Open the PR.

The magic is not the codegen.

The magic is the verification loop.

[Adam West @adamscodex]

The Robobun workflow is bigger than Bun.

It is a blueprint:

public issue
→ agent reproduction
→ failing test
→ code change
→ verified fix
→ PR
→ review bot
→ human merge

This is what “agentic software development” looks like when the product is actually testable.

[Adam West @adamscodex]

No single model stays the only answer for long

The application layer wants model optionality

[Adam West @adamscodex]

Agent-first design pattern:

1. inspect current state
2. generate proposed change
3. validate proposed change
4. dry-run impact
5. apply with permissions
6. verify result
7. log everything
8. rollback if needed

This is how software becomes safely autonomous.

[Adam West @adamscodex]

The mistake is thinking “AI-first” means conversational UI.

The real move is making your product machine-operable.

A conversational layer can sit on top.

But underneath it should be deterministic primitives:

commands, schemas, tests, logs, permissions, and recovery paths.

[Adam West @adamscodex]

Agents are weakest when success is subjective:

“page looks better”
“flow feels smoother”
“user probably understands this”

Agents are strongest when success is objective:

exit code 0
test passed
schema valid
diff approved
query returned expected rows

Verification is the moat.

[Adam West @adamscodex]

My current thesis:

The best software for AI agents will not necessarily be CLI-only.

But it will be CLI-shaped.

Explicit verbs.
Inspectable state.
Structured output.
Dry-run mode.
Logs.
Tests.
Rollback.

That’s the substrate agents can actually operate.

[Adam West @adamscodex]

Bun + Robobun is a real glimpse of agent-native software.

Issue comes in.
Agent reproduces bug.
Writes failing test.
Patches code.
Runs test.
Verifies old fails, new passes.
Opens PR.

That workflow only works because the product is testable from the terminal.