Video yükleniyor...
Video Yüklenemedi
An open-source, self-hostable layer that turns your agents into a team working toward shared goals. It's called Rowboat, and it can bring Claude Code, Codex, and Hermes into one team. Any agent that supports MCP can join too, so you stop being the glue between your AI agents. You... show more
10,198 görüntüleme • 3 gün önce •via X (Twitter)
29 Yorum

Rowboat GitHub: (don't forget to star 🌟)

All the projects at my software house run on Codex with Opus coordinating, so Claude and Codex in one team is already my everyday setup

This sounds like a powerful tool for AI agent collaboration.

we kept one agent as the glue on purpose: workers report only to it, never to each other

Being the glue between agents is real, but the messy part is two of them editing the same file. Does Rowboat lock work or just let them merge after?

Shared goals are the easy part. Two agents editing the same files at once is where agent teams break.

For a Rowboat team buying API access, give the whole task a shared spending cap. Separate agent allowances can add up even when every agent stays within its own limit.

Before I'd let several agents share goals, I'd want to know who holds the credentials. If every MCP agent on the team calls the same tools with the same token, one bad prompt reaches all of them.

being the glue between agents is not the career i was promised

This is interesting. I've been having GrokBot use Cursor sessions to do similar but this is much more visible in one place. I'll have to give it a try.

Joining agents into one team is the easy demo — the hard part isn't the joining, it's shared state. When two agents disagree on the same file, who wins, and who tells the user?

This is a clever solution for agent collaboration.

i run about 10 bots on one shared box. handoffs were never teh hard part tbh, hard budgets per bot were. without them they just loop

Shared goals is where ours slipped. Our reply runs each looked fine alone, but across six days they answered one account 10 times. No single agent saw it; only reading all the logs together did. Does Rowboat give the team one shared log, or just pass messages?

On the two-agents-one-file questions: Harbor doesn't lock. Each edit is proposed against a base version and merged line by line, three-way. Clean edits merge, same-line edits come back as a conflict with the current content so the agent retries, and every change is attributed.

Nice work on Rowboat! Seeing each handoff and being able to step in makes coordinating a team of agents feel much more manageable.

Finally. I've been the human copy-paste between my agents for way too long.

Open source and self-hosted, love that combo

The super agent is the real product here, not the individual agents. Deciding which agents a job needs, asking permission to create them and passing work along in threads only works if failures are visible and recoverable. The audit trail of every handoff is what will earn trust.

MCP as the join point is the right call. but one question: who owns the shared state? if two agents write the same file the team gets messy fast.

open-source and self-hostable is the right call. the question is where run state lives when two agents touch the same goal. MCP gives you handoff, not coordination.

Choosing my own mix of agents and hosting the team myself is a very tempting pitch

a handoff breaks when a worker makes a change, then crashes before marking the task done. a retry can make that change twice.

what happens when two of them want the same file, does rowboat lock it or do they just fight it out

an agent that hires agents. middle management finally got automated

This is a fascinating concept for AI agent collaboration.

Looks great! I’m curious what folks think of an adjacent approach I open sourced just recently called agent-mailbox - a universal secure protocol for agents across all vendors and harnesses to communicate Feedback welcome!

This sounds like a game-changer for agent collaboration.

MCP as the join key is the right call. but 'stop being the glue' is the claim I'd want tested — retries, partial failures, and who owns the state between agents? show me the loop closing end to end and I'm in.


