Hacker News new | ask | show | jobs
Show HN: Agentfab – A Distributed Agentic Platform (github.com)
3 points by razvan_maftei 61 days ago
Hi HN,

I’m the creator of agentfab, a distributed agentic platform that features task decomposition, multi-agent orchestration, model heterogeneity with custom agentic fabrics, bounded review loops, and a bespoke self-curating memory system that enable shared context.

My background is in engineering at hyperscalers where I worked extensively with foundational distributed systems. I started agentfab because I wanted an agentic coding tool that could effectively decompose and parallelize work across different model providers and agent profiles. agentfab will run locally on your machine, on your VM fleet, on your K8s cluster, or any distributed compute environment.

The clear benefit over existing agentic coding tools is that agentfab is able to fan out work across many different agents from various model providers (OpenAI, OAI-compatible, Google, Anthropic) with shared context - it’s able to break down large scope tasks into subtasks and assign the appropriate agent for each; it handles bounded review loops, artifact streaming between agents, and uses OS-level sandboxing and secure communication between agents.

You are in full control of your agent "fabric" - simply define YAML agents and specify which tools, capabilities, and special knowledge they have and the platform will know how to make use of them during task decomposition. There are 4 default agents that cover most software development tasks.

agentfab has an interactive CLI with taskgraphs and progress tracking - you can prompt agentfab to execute on end-to-end projects or chat with any of the agents in the fabric to query against their knowledge.

If you want to read more about agentfab, check out this blog: https://razvanmaftei.me/article?slug=agentfab-stateful-multi...

Check out the GitHub repo if you want to try it out. Looking forward to hearing your thoughts.

Thanks!

1 comments

Looks interesting and quite polished overall. I’m also working on an agentic workflow engine (graflow.ai) as an alternative to LangGraph, so this caught my attention. The sandboxing layer in particular looks well thought out and solid.

One question I had: in the demo, the agent builds a browser game. But that kind of task seems achievable even with vanilla Opus (plus Claude Skills / tool use), so I struggled a bit to see the core differentiation here.

Also, the project mentions "distributed" — is that mainly coordination between local processes, or does it already support cross-machine / networked execution? If it’s the former, tighter integration with remote execution (e.g. Tailscale-based peer agents or similar) could be a more meaningful differentiator. Not sure how practical that is in real deployments, but it feels like a clearer step beyond existing agent runtimes.

On the Graph-RAG part, I also couldn't fully understand how it's actually constructed internally (how the graph is built/updated and how retrieval is integrated into execution). A bit more detail on the internal mechanics there would help clarify the design.

Hi there. agentfab is distributed in the sense that it can run across different compute instances and has a control plane service managing nodes and agent instances. There's a blog post I wrote delving deeper into those capabilities with a more apt demo here: https://razvanmaftei.me/article?slug=agentfab-the-distribute...

As to your other question - agentfab uses a Conductor agent (baked into every fabric) that creates the taskgraph, which supports bounded review loops. The platform also supports pausing running tasks for user queries or change requests, which fires the decomposition step again. The end-to-end process can get a bit wordy to explain, my blog posts and the docs in the repo do a better job at that.

I'm 100% looking for collaborators on this, so if you're interested I'd love to discuss more.

Thanks for the reply — your article really helped clarify what "distributed" means in agentfab.

From what I understand, it's closer to a Kubernetes-style scheduler/worker model — so more about parallelizing computation for fully controlled systems/applications, rather than "distributed" in the sense of independent agents across boundaries.

One thought I had (just a random thought): it could be interesting to push this toward a broader notion of distribution — not quite full P2P, but something like an "agent hub", where agents owned by different organizations coordinate via MCP/A2A or similar protocols. If there were a way to manage collaborative task execution across those boundaries, that would feel pretty cutting-edge and a genuinely different class of system.