143 karma · joined November 29, 2018
We've been building agent-swarm since November last year, and we wanted to share an update on its capabilities, specially focused on the self-learning part.
After all the hype with OpenClaw, I thought that the existing architecture needed a rewrite to make it compounding. Hence, last week we implemented a self-learning core to the swarm so that it can compound.
It follows really similar ideas to the OpenClaw where there's a SOUL.md and IDENTITY.md. As it's docker based, it has some personal and shared volumes that persist, so those are used to track re-usable scripts and notes. We also added SQLite based memory that agents can write to and query. The interesting part about it is that there's personal and shared memory, which allows the lead to propagate learnings across the swarm!
We've been using it non-stop for the last week, and I already see the compounding effects. E.g. we have a morning scheduled task that makes the lead assess the previous day work, and figure out ways to improve it's processes, and it got better!
To end, note that it's fully OSS and it's as easy as deploying a docker compose to a VPS, or even locally. It's core is based on an MCP that the lead and all workers share, which allows you to impersonate the lead locally to control the swarm from your coding agent too!
We implemented a super simple UI at app.agent-swarm.dev that runs in the browser only so you can put your API url and key to see it in action.
P.S.: It uses the claude CLI only now, so there should be no issue with the Anthropic terms, and it's really thought to be self-hostable.
P.S.2: Obviously, all the agent swarm code has been written at 95% by agent swarm via Slack :D
If you have doubts or questions about the architecture, or what we are planning to build next, happy to chat in the comments section!
- https://github.com/desplega-ai/qa-use CLI to control browsers and tests locally, coding agent friendly - https://github.com/desplega-ai/agent-swarm OSS agent swarm that can be deployed with docker compose using Claude - https://desplega.ai platform that supports qa-use backend
Cheers!
No auth, just vote and it will store that. Note that you might vote more than once.
Code at https://github.com/desplega-ai/ai-toolbox/tree/main/arewedoo...
Conjecture: Embrace boring tech, to build new tech.
We built an MCP server that coordinates multiple Claude Code instances running in Docker containers. It follows a lead/workers pattern, and the MCP glues the agents together.
They get a full ubuntu with sudo (YOLO) and a way to spawn services with PM2 (so they are persisted and restarted on container restarts).
Also, they have persistent workspaces, both for personal and shared across the whole swarm, so the lead can micromanage the workers more easily :D
It's super easy to setup and deploy. You can run it locally, really the only thing required is running `claude setup-token` and good to go.
WIP:
- Only tested with Claude Code, though architecture should support other CLI agents (Gemini CLI, Codex, Opencode). - Horizontal scaling works but coordination gets chatty with many workers (needs prompting)
Open to feedback on the architecture, or similar projects!
1. Relevance to HN (e.g. OSS, relevant tech articles, etc.) 2. Similar posts matching (e.g. if it's repeated, or low-impact ones) 3. Provides recommendations on title, description
To be fair, it's pretty "dumb" now, as it does a mix of deterministic queries to the database + LLM analysis to build the report.
You can see the plan for the idea tester here https://github.com/desplega-ai/ai-toolbox/blob/main/willitfr...
Any plan to make it oss?
It was pretty straight forward to do, specially thanks that all vibe coding apps have iframes with the dev sites, so it's easy to get the url. Also, funny enough, they are generally served with a token param, that is perfect to use externally as in our case: to vibe e2e test the app.
But... obviously it might get flaky if they change something, probably need to vibe Qa the vibe QA :D
Repo for the extension here https://github.com/desplega-ai/vibe-qa-extension
We focus on making it as fast as possible, integrated into CI, MCP for local dev, and support both an autonomous (we call it discovery) and guided test creation approach.
We believe that in the era of vibe-coding, quality is key, as we are lazer focus on building a solution that scales with your product, and removes the burden of QA from your team.
Technically, we built an in-house engine that is in charge of generating the tests, that speeds up and gets better the more you use it.
tldr: Context control is key for complex flows, if you are not doing that you are just guessing.
What do you think?
Also part of our format is specially finding deterministic way of running steps, with automatic healing when failed. And we also build the whole system in a way that is self-hostable, so in the cases you mention you could be able to have control over what is run and where.
Also about the CLI thing you mention, we had cases where people did not use MCP clients, so I actually wanted to expose a way to interact with the tools directly without the need of the server running!
If your challenge is flakiness, this should help initially. Unfortunately, there’s a lot of work in our engine, and a custom system to handle operations that goes beyond vanilla Playwright so running it locally would be quite challenging.
Funny enough, one of the most challenging things while building it was to be able to remotely control the browser that runs locally (been using https://localtunnel.me/ for it), while making sure that it does not impact the user too much.
Also, I wondered if there's someone that is shipping CLIs with an "mcp" command, as seems that having a normal CLI for the funcionatilitie sof the MCP would make a lot of sense + with the option of running it as an MCP if the user wants to.
Have you seen this pattern?
Also, as there's a lot of buzzwording around MCPs, any of you that been using an MCP as a daily driver? For me it was the github one, specially for code search and stuff like that.
I'm building a company and as we are early stage, we have a newsletter for people that might be interested in our progress, and one of them is a good friend of mine. Nevertheless, he asked to be removed from our newsletter, and instead of just unsubscribing like a normal person, he said "just give me an RSS feed"... So I built this while having lunch today...
The interesting technical bit: Resend's broadcast API doesn't return email content (it's documented but still surprising). So you can't just fetch broadcasts and serve them as RSS...
The workaround is to:
1. Fetch broadcast metadata (subject, from address) 2. Fetch all sent emails separately 3. Match them by comparing subject + from fields 4. Cache everything in Vercel blob storage
The cron job respects Resend's 2 req/sec rate limit, so first sync takes ~4 minutes for 370 emails. But subsequent runs are incremental - only 9 seconds when there's nothing new.
Built with Next.js API routes + Vercel Blob + Vercel Cron. Works surprisingly well for such a hacky approach. The matching-by-subject thing feels fragile but hasn't failed yet jeje
---
PS: Are you still using RSS?
PS2: I initially thought of using Google Groups for this, but seems it's not supported. Sad :/
PS3: Thanks Claude Code 2 <3
When I worked in a bigger company in the past we used postgres to store raw blobs for data transformation pipelines (+1T table), and queries to it were instant due to good index usage. Others might have gone to BigQuery or something else making more complex and def more expensive.
PS: I love tools like Temporal, Hatchet or Oban that rely on Postgres and can scale a ton too.
As we all know, in large orgs the form issue has been solved by using trusted components with unit tests. What worries me is how somebody were actually to provide valuable bugs, those that keep you up at night trying to fix them.
I worked with exploratory testing companies that would report hundreds of bug that nobody ever cared about. Does YC really care about the bug that they found, or have a business impact?
(Actually coming from building a company in this area now https://desplega.ai, you can try it out for free too)
I just received the email from Github that they enabled Spark for my account, and I wanted to give it a try. I asked the same prompt to Spark and v0 and Spark did pretty good, and actually it let change the theme (in an opinionated) way that is pretty easy if you want to try some presets.
It's a simple text to JSON tool using a few common models, using an API key that is stored in the browser, you can see it in the repo here: https://github.com/spark/tarasyarema/text-to-json-schema. Surprisingly I always end up doing a lot of this type of operations, and wanted to give it a try, maybe I end up using it :D
What do you think about Github moving into this realm of generative apps? Will they win the lovables' and v0s?
Only bad thing is
1. You can not make the link public... you need to be signed in into Github... 2. Tried to deploy with Vercel and seems the spark module is private...
Also in the first try it used a function spark.llm which was not even exported which felt weird, and actually required no API key... I could not find any docs around it, anyone has some insights on it?
Cheers!
I always wanted to contribute to OSS. I tried direct ways to do it, like Hacktoberfest, finding good-first-issues, reading contribution guides. Nevertheless, after a while I found that the best way to do it (for me) was to actually build something I was passionate about first, which naturally led to contributions.
tl;dr: Stop trying to force contributions. Start by solving your own problems, and the contributions will follow.
Curious about how did you started to contribute? Any fun stories around that?