learning the layer below your tools changes what you can understand.
@unclebobmartin tells @mattpocockuk why he used to suggest a weekend of assembly language, in their conversation about agents.
if every change to a house cost $1, would you plan it the same way?
@unclebobmartin uses that thought experiment with @mattpocockuk to explain his preference for small changes and feedback.
keep the quality values. reconsider which human work habits an agent needs.
@unclebobmartin makes that distinction in his chat with @mattpocockuk.
a module with a clear purpose gives an agent less conflicting context to deal with.
@unclebobmartin and @mattpocockuk discuss why clean boundaries still help when agents write the code.
architecture rules can be checked after the agent writes code.
@unclebobmartin explains his dependency checks to @mattpocockuk: specify which modules may depend on each other, then require violations to be fixed.
writing the feature is one stage of @unclebobmartin's workflow.
in his chat with @mattpocockuk, he walks through separate agents for code cleanup, mutation testing, and qa.
a qa procedure should actually use the system.
@unclebobmartin tells @mattpocockuk how he turns a human-written story into acceptance tests and a procedure that works through the ui.
focused tasks help keep agent context manageable.
@unclebobmartin's chat with @mattpocockuk covers fresh contexts, parallel work, and the startup cost of handing work between agents.
give an agent a check it can run, then let it work until the check passes.
@unclebobmartin explains the quality tradeoff to @mattpocockuk. that extra work has a cost too.
a longer instruction file doesn't guarantee better results.
@unclebobmartin talks with @mattpocockuk about context getting lost in the middle, and why he keeps the initial prompt short.
messy code can trap an agent in a cycle of fixing one bug and creating another.
@unclebobmartin describes seeing this happen in his chat with @mattpocockuk. cleanup still matters when an agent writes the code.
a passing test suite can still miss broken behavior.
@unclebobmartin explains mutation testing in this chat with @mattpocockuk: change the code on purpose and check that the tests notice.
product work means understanding enough problems to find a solution that addresses many of them. @thdxr explains to @GergelyOrosz why that still takes experience and careful thinking.
agents make it easier to replace old patterns across a codebase. @thdxr tells @GergelyOrosz why opencode is using that ability for cleanup and a better place to work long term.
ai time savings don't automatically become more output. @thdxr tells @GergelyOrosz why people may finish their work sooner, and why motivation and compensation change that outcome.
Anthony Cueva retweeted
Banks, funds and fintechs are on stage at @aiDotEngineer NYC this week, talking about agents in production.
One question we'd add to every deployment checklist:
When the agent is wrong, can you show what it remembered, where that came from, and who corrected it?
gpu demand can outpace the supply needed to serve it. @thdxr tells @GergelyOrosz why opencode has to reserve capacity and pay ahead for inference.
Anthony Cueva retweeted
¿Firmas o negocias? ¿Riesgo o seguridad? ¿Descanso o pastilla?
Un juego gratis de 4 min donde eres la webcam de un dev en Lima y decides tú. Ideal para juniors y estudiantes.
binnicordova.github.io/softw…
Ayúdame a ganar un cupo en #HackTheAndes: juega, comenta tu final y RT 🙏
