Gene Taylor

Your sprint run on my team's code. Engineering Manager for Jira backend @Atlassian

[Gene Taylor @SplicedGene]

@GergelyOrosz Code review as a quality gate is going to fade away, but until we commit to full dark factory software we will need to introduce a new ritual to replace the socialising understanding aspects of code review

[Gene Taylor @SplicedGene]

@pav_light I'm a big fan of Rands meeting taxonomy - the useful ones are either for alignment or creation. You need know what kind of meeting it is ahead of time, who the ref is and what the agenda is

[Gene Taylor @SplicedGene]

This man gets it

[dex @dexhorthy]

stop doing loops, start doing graphs

[Gene Taylor @SplicedGene]

@n1tr0maverick Next up is someone training a transformer on vintage hardware

[Gene Taylor @SplicedGene]

@Noahpinion Why change it? With Fable and 5.6 they are finally fun again to talk to

[Gene Taylor @SplicedGene]

@koomen @paulcapriolo It’s a fun UI, but it reinforces my theory: given enough time, every workflow evolves into a Kanban board.

[Gene Taylor @SplicedGene]

@oscrhong School can work because of the forced accountability.
Look at MOOCs: they failed because there was no social pressure to push people through the hard parts.
The most successful 'gym for the mind' is going to have the good kind of peer pressure.

[Gene Taylor @SplicedGene]

@dannolan You have to hit them with the expression

[Gene Taylor @SplicedGene]

@GergelyOrosz I'm really interested in seeing how much we are going to be able to more explicitly borrow from manufacturing quality control practices - poka-yoke, acceptance sampling etc

[Gene Taylor @SplicedGene]

@oscrhong The episode for anyone that didn't catch it at the time https://t.co/46MlWP2g6U

[Gene Taylor @SplicedGene]

@oscrhong Karpathy was making a similar point on the Dwarkesh Podcast - we are going to need to develop something that can motivate hard cognitive in a world where it's not necessarily required to generate results.

[Gene Taylor @SplicedGene]

This take has aged very well - a significant amount of AI anxiety vs AI excitement is downstream of this cultural divide

[Kent C. Dodds 🐨 @kentcdodds]

This take by @zachlloydtweets is more relevant now than it was in 2022: https://t.co/sJxpHwCFGF

At this point (not to sound alarmist but...) I would say the code-first developer is an endangered species 💀

[Gene Taylor @SplicedGene]

@simonw Still using mine - as a personal assistant. I'm not super opinionated about the harness itself, the value is mostly in having an always on agent that can be a routing layer over deterministic scripts

[Gene Taylor @SplicedGene]

@wabzqem It's Cmux + harness TUI for me.
More than anything I want a container around each agent session and it's task context to reduce the cognitive load of having multiple agents on the go.

[Gene Taylor @SplicedGene]

@falgunichitkara Yes, being able to fork an agent for tasks like this is a undersold productivity unlock. It completely removes the tension between carrying an open loop or breaking your flow.

[Gene Taylor @SplicedGene]

The Generator–Validator pattern is simple but powerful, and the tokens spent on the adversarial pass more than pay for themselves in output quality.
Curious where else people are finding value from it. 4/4

[Gene Taylor @SplicedGene]

Everyone is trying to use AI to clear the bug backlog. The problem isn't having it triage bugs, it's that the AI confidently files garbage bug reports.
Here's a simple trick that fixed that for me: one Triage agent that proposes tickets, and a second adversarial Validator agent that only sees the draft + codebase and tries to tear it apart. 1/4

[Gene Taylor @SplicedGene]

The results: from 93 starting bugs I ended up with 29 ready to move to my team's development backlog in Jira, 22 flagged for a human to review, and 5 false positives rejected before any time was spent chasing a fix in the wrong place. 3/4

[Gene Taylor @SplicedGene]

@davidcrawshaw Having a second agent with fresh context review the work of the first is so simple and yet so effective. I have a lot of success with this beyond code review when getting agents to do codebase analysis