Just created a PR from a github issue via a serverless background agent, and it's entirely open-source at humanlayer we have been thinking a lot about what good open-source software factory building blocks look like to push on this I built something I've wanted for a long time: a hands-outside-the-box approach of running a lightweight agent in a runtime with a virtual fs, and only using a container when necessary - faster, cheaper, etc combined with really good webhook / event bindings for GitHub, slack, and linear we built two components for it: 1. - effect-native webhook subscriptions for slack/linear/github; like the @vercel chat SDK but effect + more opinionated about software factories; handles threading & debouncing event deliveries and stuff like that 2. mentions & other events spin up our Fold harness ( think kinda like @pidotdev in effect; fully isomorphic, runs anywhere) infra is built on @cloudflare workers and durable objects via @alchemy_run - each thread/channel/issue/pr gets its own mailbox, subject to delivery policies (queue, debounce, batch) mailbox messages wake a fold agent session running in a Durable object, without a sandbox to start it uses cloudflare/computer for a virtual DO-backed filesystem for each issue/PR's git branch the agent can run most commands it wants against the DO's virtual filesystem and the agent can boot a container with the FS mounted into it via FUSE only when it decides it needs one to run heavier commands tools are generated from the webhook event so the agent gets tightly-scoped tools to e.g. create a pr from the issue it's on, or to list failing checks only from the PR it's bound to; plus it's easy to interleave deterministic code it has live progress updates, can react to issues in its thread, etc (linear + slack modules support their respective streaming agent activity & agent plan APIs; for github we fudge it a bit by editing issues in live time) it even has an API so agent work can be handed off to another system (e.g. a k8s cluster that has super thick sandboxes if you need > standard-4 sized-containers from cloudfalre) and that system can post back streaming updates & update the status of the agent run, so it's pretty de-coupled it's wired to automatically investigate CI failures and some other neat stuff open-source at a good example of it in action:

if you're building an agent or a harness, you need to be an order of magnitude closer to the code than you would be normally models are VERY bad at building harnesses & agents - not sure if it's lack of training data or smth else but they have exceptionally bad intuitions about context management and the agent loop and caching and a lot of other components that matter a LOT it is very high-leverage to design the code yourself - if not typing it by hand, still do the program design yourself

man every time I give "adversarial review" another try it just comes up with a bunch of defects or deviations from the plan or design doc that aren't actually deviations it turns out when you tell an LLM "find problems" it finds them, even if they're not real blindly trusting in review loops to ensure adherence to a design doc is gonna create a TON of extra code that does nothing for you

it's funny how I hated my ARM assembly class in college (to be fair the instructor and TA's were quite bad) but i look back on it very fondly i really miss writing code by hand :( writing quicksort by hand from scratch seemed awful at the time but being so close to the control flow and memory was amazing

crazy that locale doesn't do some super lightweight inference to guess what I would like each week - could even do this with some basic clustering or multivariate analysis without LLMs the recommendations are not great but we have a super consistent pattern of what we order the recommendations are never good

lots of coding agent memory systems are unhelpful - at best i have said this a lot and some folks think I'm against agent memory in entirety which is not true - I just think that users should be on the loop that generates memories and should be able to enable / disable the system or individual memories at will and should be able to dismiss disruptive or unhelpful memories putting a user on the loop also lets you build a better data flywheel over time to improve the quality of recommendations live in @humanlayer_dev btw lmk if you want beta access

so @jhleath doesn't want you to know this but you can just mount an archil disk to your machine and configure your system to use the mount point as additional swap space allowing you to download more ram on S3

the superintelligent bash homunculus remains undefeated at chrome profiling & performance optimization the only downside is my machine feels haunted with all the windows popping open and closed