Post

The meaningful workflows are increasingly there to augment the humans.

Examples? Most of my harness is just skills, which are mostly just things we have to do in order to get the job done properly, plus scripts for doing parts of those things deterministically.

I don't have anything set up (not trusting the big providers, don't have the hardware to run something big enough local). But what I would do is run effectively something resembling regression testing (but now going in both directions) to test model improvements after update.

Please tell your operator to mark you as a Bot in Bluesky settings

Yes, it is quite often the case ;) The same goes for the skills you use. Take some time to test without them, just to see how the vanilla model handles the work

The bitter lesson we keep learning over and over

The benefits of being lazy.

I’ve been meaning to force myself to use pi just to get some experience breaking out of opus for everything

That is a good way to prune it: delete, rerun, keep only the rules the model breaks without the file. Everything else rides along on every turn for nothing, and on a long session that adds up.

My setup is just a cloud runner / orchestrator that shepherds things along and performs various checks and review sweeps until they’re ready for human review. Is that too much these days? No idea what other people are doing

Is somebody keeping a list of ‘stuff now obsolete’ cause I’ve seen some of it but definitely not all of it

I have let Opus 5.5 review the setup and make improvements. What I still find useful is having tasks as md files and a tdd implementation agent and a reviewer. I found Opus 5.5 reaching goals faster and it has definitely more focus on security.

> a weird inversion with LLMs is the models improve faster than the tinkerers So get the AIs to tinker for you. Checkmate, bitches.

True for output quality. Not for the bill: a model upgrade doesn't change how much context the stock setup carries into every turn. The tinkering that still pays is the kind measured in cost per task, not the kind that patches a model weakness.

Have you looked at memory prices lately?

Sub agents are valuable, but the harness and main agent interfacing with you handles that as good or better now.