kerrow

software, self improvement, serendipity.

[kerrow @itskerrow]

happy 4th https://t.co/o6Dr4AnJYP

[kerrow @itskerrow]

@kenwheeler yes https://t.co/WrFwctHkFi

[kerrow @itskerrow]

@ghumare64 yes - an agent loop isn't much different than any other loop. i created https://t.co/0W4CwmLqHm to embed in my python projects and not be vendor-locked to anthropic with claude agent sdk, while still having a sophisticated harness

[kerrow @itskerrow]

i use https://t.co/nMp4GZYjWP - i can tell my agent to just orchestrate my swdev kit flows which include a few different types of planning and implement flows. i have a high level working document, and it delegates the detail work to dedicated fresh-context agents and oversees, coordinates, and reports on their work.

[kerrow @itskerrow]

stepwise 0.44: settings page redesigned by a council of 4 models. coordination validator with 350k fuzzed flows / 0 soundness bugs. kits for flow organization + registry sharing. cron/poll scheduling. fixed silent output corruption when concurrent jobs share an agent process.

https://t.co/S896HfqtE0

[kerrow @itskerrow]

tests at exist at multiple layers for different reasons and have their own lifecycles. tightly coupled tests that catch inadvertent and subtle interface contract changes need to be modified or thrown away when that interface changes, which happens in the real world where most interfaces have few consumers and change is an option. move fast-break things-just not on accident.

[kerrow @itskerrow]

session orchestration and context management is not totally sorted out yet. there are useful patterns that can't be expressed with subagents - even when recursion is allowed

[Aakash Gupta @aakashgupta]

The fact that every coding CLI just independently arrived at the exact same architecture tells you what the real bottleneck in AI coding actually is.

Claude Code shipped subagents in July 2025. Gemini CLI is shipping them today. The agent definition files are nearly identical: markdown with YAML frontmatter, isolated context windows, hub-and-spoke delegation, tool restrictions per agent. The convergence is so tight you could swap config files between tools and they'd almost parse.

This happens when multiple teams solve the same underlying physics problem. And the physics problem here is context window management.

A 200K token window sounds massive until you're three hours into a real engineering session. Every file read, every grep output, every test log stays in memory. The model starts attending to noise from two hours ago that has nothing to do with the current task. Response quality degrades because the context is full of garbage.

Subagents are garbage collection for AI. Ship a task to an isolated context, let it burn through 50K tokens of file reads and searches, receive a 500-token summary back. The noise never touches your main session.

The competition in coding agents for the next 12 months is about who builds the best context orchestration layer. The model is approaching commodity. The memory architecture is the moat.

[kerrow @itskerrow]

@MichaelMammoth @systemdesignone ... and our commit history goes back to 2009 when we switched off svn

[kerrow @itskerrow]

@MichaelMammoth @systemdesignone i run a company with multiple teams of engineers, am an engineer myself as well, and have tenures exceeding 10y of numerous staff

[kerrow @itskerrow]

@asaio87 there is plenty of fear (and it's not unsubstantiated) among devs. dario said in his dwarkesh interview in feb that he expects 90% of swe jobs to ultimately go away...

[kerrow @itskerrow]

coevolution is the only long term solution to alignment.

[Bojan Tunguz @tunguz]

I honestly believe that the best case scenario for human future is that we become symbiotes with the machine intelligence.

[kerrow @itskerrow]

lean into the attention mechanism. unlock intelligence with focus

[kerrow @itskerrow]

@nateberkopec they are happy to delete things if you ask them to

[kerrow @itskerrow]

@pmddomingos *some

[kerrow @itskerrow]

@tekbog i'm still waiting for the gpt store!

[kerrow @itskerrow]

@adamhjk every generative agent needs an adversarial agent.