OpenHands co-founder Graham Neubig requests feedback on a proposed tutorial to build an open-source alternative to Claude Tag

@gneubig looking forward to this!!

@gneubig My infra colleagues built something similar and shared about some of the key features here which might serve as inspiration

@EntropyFighters Amazing, this is a really great list, and I think OpenHands automations support most (but not all) of these things already. I'll add this to the tutorial

@gneubig I'd say for me: - per-channel isolation (code agent works on codebase, marketing agent on google docs) - orchestration from Slack but ability to use other integration (in particular GDocs / GSheets / Gslides) - how to do the same on Discord instead of Slack (more OSS friendly)

@gneubig Out of curiosity, when did OpenHands add Slack integration? Am I remembering correctly that it was present back in 2024?

@gneubig Just another open-source clone that’ll be forgotten by next month.

Cool — here's how I handle the two 👇

Budget ceiling w/ graceful degradation: every run gets a hard cap (tokens or $). The trick isn't the cap, it's what happens at it — instead of hard-dying mid-task, the agent checkpoints: writes where it got to, what's left, and a "resume from here" note. Hitting the ceiling becomes a pause, not a lost run.

Glad it's useful. If I had to bet on the "not all," it's the budget ceiling with graceful degradation (checkpoint + yield at the cap, not hard-die mid-task) and the act-vs-ask escalation rule — those are the two almost everyone skips, and they're exactly what flips it from "cool demo" to "I'll leave it running unattended." Happy to share how I handle them if it's useful for the tutorial.

@kentyman Yep, although tbf that was not open source.

We now have better support for open-source automations and I'd like to use that to set it up!

https://docs.openhands.dev/openhands/usage/automations/overview

@EntropyFighters Yeah definitely useful!

Act-vs-ask escalation: an explicit rule for when the agent proceeds vs stops and asks. I gate on two axes — reversibility (can I undo this cheaply?) and blast radius (what does it touch?). Low + low → act and log it. High on either → checkpoint and ask. It lives in config, not vibes, so it's auditable.

@gneubig Sure, but can you make it shift paradigms?

@_lewtun @gneubig https://vercel.com/eve

Eve by @vercel is quite nice and supports the different harnesses / models / connectors (like slack)

@_lewtun Very cool! Thanks for sharing.

@gsarti_ Great list thanks!

@ChainZenit @gneubig Jeez. Even if it was the case, so? Besides Graham did say he wanted to make a tutorial which would prove useful regardless of anyone ever uses the project.

Both share a spine: an agent you leave running unattended needs defined behavior at its edges — out of budget, out of confidence, out of scope. Capability gets you the demo; deciding what happens at the boundary is what makes it something you trust. That's the part worth putting in the tutorial.