In an afternoon, I was able—with the help of OpenHands—to eliminate much of eight years of technical debt in two medium-sized Projects. I had been deferring this maintenance for a couple of years and was dreading the tedium of refactoring and modernizing the code. OpenHands made it a much more pleasant experience.

While only an internal implementation, was able to quickly replace Claude Agents SDK, which relies completely on Claude Code, with OpenHands SDK for an elasticsearch search agent in like a day's work. This is the first step towards my company being able to run our agents on locally hosted LLMs that can be run on prem! Thanks! PS: Because of how thinking blocks work in anthropic, swapping them with your think tool actually made the agent go faster! They don't stall those tokens

🚀 We want to spotlight community voices. Got a quick 1–2 sentence quote about how OpenHands helped you? Drop it here or use this form: https://openhands.dev/quote We will attribute the quote to you and select quotes get a limited-edition swag. Deadline: Nov 10. Slack Conversation

Just wanted to share something i've built using openhands sdk: https://github.com/mr-karan/hodor Have been using this internally in our org's Gitlab CI pipelines and the results are amazing. openhands is pretty good at writing the agentic loop declaratively and has the right set of components to build your own agents easily. 🙂

I installed that OpenHands agent in the Zed editor via our new ACP (Agent Client Protocol) support, and then installed the Tavily Search MCP. Now I can ask any question I want in my editor sidebar, and it will answer me, including Zed tech support questions 🙂

I asked our codereview-roasted agent to review a grant proposal that I’m working on for CMU 😭

If you hook up OpenHands to the nanobanana mcp you can generate images and diagrams 🙂 It’s a lot of fun

Tokens: ↑ input 3.57M • cache hit 0.00% • ↓ output 21.68K • $ 0.00

I have no clue what happened in my self hosted setup but I am now able to have very long running conversations and finally got to witness the context compression in action. Using a full sized qwen3 coder 480B model. I'm able to make really good documentation for new codebases. I don't think the model has actually processed 3.57 million tokens in this chat, but regardless, a week ago things would break after just a few tool calls. Would love to get an accurate metric of the context length in my requests, but regardless, I'm very VERY happy right now.

This prompt worked surprisingly well

in https://github.com/OpenHands/typescript-client, list all the open dependabot PRs. Use the GitHub API to merge lastest changes from main into those branches. Once CI/CD is passing, merge the PRs

can Codex/ClaudeCode use this as well?

I'm excited for the rollout of the verifiers stack! I recently discussed the "death of SWE-Bench Verified" (due to contamination) in a segment of our AI-First Show: One of my students commented with a couple of questions, one of them about agent benchmarking, and I just replied referencing the OpenHands trajectory-level verifier. Aside from ensuring coding agent reliability and increasing trust, I think this could also be an important verification component in agent benchmarking.

Eu acabei de chegar. Perdi alguma coisa?

oohyyeah The all-hands-bot is the smoothest PR experience ever 🤩

Not yet a "success story" but a work in progress: my interest in agent harness orchestration has culminated in an article (that I'm pitching at the moment), and a 😆 And of course, my own orchestration system I'm working on. I'm loving the new OpenAI Symphony system (driven from Linear issues and GitHub PRs), and I'm using it to create OpenSymphony with several improvements. It will include a Rust TUI orchestrator dispatching work to OpenHands via

agent-server

. I'm using the reference OpenAI Symphony implementation (which uses codex app-server) to do the new implementation. I just started that build today. We'll see where it's as after a few hundred million more tokens! I already had

chatgpt-codex-connector

wired up to do automated code reviews, and it's so fun to see the system automatically responding to them and addressing them. In future iterations I can explore integrating it with emerging OpenHands work on verification stack and PR reviews.

I trained OpenHands to shop for groceries for me: https://github.com/neubig/workflow/blob/main/skills/giant-eagle/SKILL.md

I’m getting OpenHands to create a React version of our company web site, and with Claude 4.7 it has now been working about 7 hours straight 🙂

Using fresh openhands conversations as subagents to recover from missing github token

Wanted to share out how you can orchestrate multiple agents / harnesses (like Claude Code) all controlled by OpenHands - https://github.com/rajshah4/openhands-multi-agent-demo (combine this with an something like an iterate skill - https://github.com/OpenHands/extensions/tree/main/skills/iterate and go crazy and burn lots of tokens).

I used openhands, not for code this time, but to do historical research to a Resistance Group: https://github.com/sjorspa/hembrug-verzet

I love love love switching between LLMs in the same conversation! amazing ❤️ clapping

Fun! OpenHands agent is sneaking in

Open WebUI

, a popular “speak to your LLM” interface (140k+ stars)

Watching the Watchers 🫠 Context: openhands posted a video in a PR, of its own testing of a UI feature (cc: @average-processor-75195 your agent’s QA video is cool ❤️) rubber stamp

Hey folks, I put together a small experiment: OpenHands OpenWiki AutoDocs - github.com/rajshah4/openhands-autodocs The idea is to use an OpenHands skill + Agent Canvas/automation support to generate and maintain lightweight repo docs under

openwiki/

. It’s inspired by LangChain’s OpenWiki project, but supports additional documentation tools like GitNexus, and relies on OpenHands Automations: • Cron/event-based updates • GitHub integrations for PR-driven docs refreshes The repo includes skills for running either OpenWiki or Gitnexus as well as some example results. (We could expand this to other documentation tools / integrations as well) Would love feedback on the approach, especially around docs quality, automation patterns, and how this could fit into repo-maintenance / SDLC workflows. I have talked to quite a few teams looking for an automated documentation solution for agents and humans and this could be a good starting point.

🚀 *I think I just had my "holy sh*t" moment.* Repository: github.com/tamzrod/KDSE For the longest time, AI-assisted development looked like this: > Prompt → Code → Debug → Prompt Again Today I tried something completely different. Instead of asking the AI to write code... I asked it to engineer. Using my experimental methodology, KDSE (Knowledge-Driven Software Engineering), the workflow became:

Knowledge

↓

Architecture

↓

Implementation

↓

Verification

Then I gave it ONE engineering request. > "Build an Industrial Power Plant Monitoring Dashboard." That's it. No architecture prompt. No implementation prompt. No debugging prompt. No screenshot prompt. The runtime handled the engineering process. The result? ✅ Knowledge Base ✅ Architecture ✅ Backend ✅ Frontend ✅ Apache ECharts Dashboard ✅ Mock Telemetry Simulator ✅ Automated Verification ✅ Playwright Screenshots 🤯 From a single engineering request. This isn't about generating code anymore. It's about giving AI an engineering process to follow. I honestly think we're moving beyond prompt engineering. We're entering engineering runtime territory. This is still experimental, but today was the first time I felt like... > "I can build software by engineering the knowledge—not by micromanaging the code." I'm ridiculously excited to see where this goes. 🔥 #AI #SoftwareEngineering #OpenHands #LLM #ApacheECharts #KDSE

https://openhands-ai.slack.com/archives/C06U8UTKSAD/p1784030756385879?thread_ts=1783608548.448879&cid=C06U8UTKSAD I had some success so why not crosspost in success-stories!

I guess everyone is doing this already? 😅 Yesterday I finally changed my browser from Dia to Comet. I was dreading the 226 open tabs "loss" - because of course! So I asked my agent to read and re-open them all in Comet. And since we were at it, I also wanted it to pick up all my X bookmarks and all Dia bookmarks, read and reorganize them. It just did it. Over 1,500, nice and tidy. To note, it was using its own browser, not X API. ✅

I make use of OpenHands Vulnerability Fixer as part of the toolkit I recommend in an AI Cybersecurity Playbook I published to leverage Kimi K3: trilogyai.substack.com/p/ai-cybersecurity-playbook-kimi-k3

Saw fireworks used openhands with K3 (although they had to make some changes) - https://x.com/fireworksai_hq/status/2084374420295209209?s=46

I've been pushing Agent Canvas pretty hard lately, vibecoding some personal projects I find the current experience sort of frustrating. I'll have 17 ideas for fixes, but managing 17 agents is too much. I can throw them all at the same convo, but the agent gets confused. Some of the ideas are big, some are small. Some of the ideas touch similar parts of the codebase and need to be sequenced; others can be easily parallelized. Some ideas should get Fable, others should get Sonnet. Also annoying to have to watch each convo, wait til it finishes, see what's been done, and say "push to main" or "open a PR" over and over again. So I built a Manager! I have a UI where I can rapidly submit tickets. There's a 1m cron automation that checks the board and kicks off conversations as necessary. It will funnel similar tickets into the same convo, and knows to hold off on tickets that will conflict with WIP. It also knows to keep at most N conversations going at once. This feels way nicer--I'm not really babysitting conversations anymore, just firing off tickets and then clicking "verified" once I've confirmed the feature has landed

FYI: These folks in Australia have been blogging about OpenHands wallabytoken.com/blog/p/openhands-memory-remote-backend