[ale]

I don’t know how to phrase this without sounding like an arrogant idiot but seriously: what are people actually programming with agents + worktrees + harnesses + tasks + skills + whatnot? Most workflows I see people adopt involve large amounts of infrastructural fluff only to (more) quickly generate what I (anecdotally) have seen is somewhere between code generation of boilerplatish React/laravel/your-fav-framework components for web or native, and niche toy apps for mostly personal use. My very limited usage of agents has been for scanning large (bloated) codebases to get rid of unused code, meaning time consuming and tedious tasks. But it seems the general trend is that programmers just want faster horses?

[sowbug]

You could zoom out a bit and rephrase the question.

Your great-aunt Ida died and left you a consulting team of ten pretty good software engineers. The team's contracts all just ended, so starting tomorrow they'll be idle. Ida said you must run the business for at least two years (fortunately, overhead is already paid for), or forfeit your share of the inheritance. After that you can keep going or liquidate it.

What do you do?

[KurSix]

People are just playing around with parallel agents because it looks cool on Twitter. In real prod, 90% of your time isn't spent typing lines of code, it's spent trying to figure out implicit business requirements and debugging undocumented legacy spaghetti. Agents sitting in isolated worktrees are completely useless here - they'll just rapidly and in parallel write code that completely fails to solve the actual business problem

[SkyPuncher]

You’re seeing that primarily because it’s what people can show off easily. Side projects they do for fun.

I use all of this stuff daily at work. Normally, I’m working on 2 to 4 features in parallel (so worktrees). This might not be simultaneously, but it’s at least across days or weeks.

Skills, agents, tasks, etc are really about creating repeatability in certain parts of my workflow without needing to be hands on.

[sam0x17]

Yeah perfect example, the main thing I _would_ use multiple agents on is optimizing/benchmarking code, but for that you specifically can't use worktree, you need one agent per machine or they'll taint each other's benchmarks

[KronisLV]

Some changes within pre-existing codebases (oh hey, we need feature or mechanism X).

An entirely new internal system with LLM code review, DB migration tracking, time tracking, standups and Teams integration.

A new system that trains neural nets to recognize crops based on Sentinel-2 satellite data (the neural net works okay, mowing and ploughing is harder with a mostly heuristic approach since I don't really have labels).

A new system to migrate somewhere between 1000-2000 forms between proprietary solutions where a team of people have spent a year with limited progress, whereas I'm generating the codegen tool that does most of the work, with the remainder being left up to AI.

A new project linting tool with Go + goja to allow writing rules for validating project stuff in ECMAScript, a bit like ESLint just stack agnostic and can be deployed as a 10 MB executable, to also control stuff like architecture and project conventions that the other tools aren't really geared towards.

Also wrote an OpenAI/Ollama/Claude proxy that allows using on-prem models running on another server through Ollama/llama.cpp and also using AWS Bedrock models when permissions are configured.

Also a bunch of Ansible configuration for stuff like a self-hosted Sentry instance, debugging that piece of shit would be so hard and annoying without something that I can throw logs at (because for some reason they think that having 70 containers running for what should amount to one piece of software is okay).

Also wrote a personal tool that lets me use VLM and Whisper and PySceneDetect and some other stuff to produce EDL so I can take a 3 hour long video and cut it down to 1 hour with LLMs using the transcripts/timestamps (aligned with words, so not too many awkward cuts) that I can then import into DaVinci Resolve for further editing.

Also migrated the apps I host from Contabo VPSes to Proxmox VMs (Hetzner dedicated server from the auction) and went from Docker Swarm + Portainer to pure Docker Compose, also moved from Drone CI to Woodpecker CI and also got rid of the old deprecated Bitnami container images.

Also migrated my homepage from an ancient Ruby and Rails version to more modern ones.

Also wrote a few scripts to replace YOURLS with just an Apache install, the config for which I can automatically append new shortened links to.

I don't even need worktrees or custom skills for most of this, just Claude Code and a subscription, since paying per token would make me go broke.