Show HN: Quibbler – A critic for your coding agent that learns what you want
github.comIn other words, the Quibbler siphons wrackspurts away from your code.
There's so many agents to handle my agents, I'm gonna need agents for my agent agents soon.
Reminds of factory factory factory! https://factoryfactoryfactory.net/
More explanation here that I found by Googling around. Though not sure it has more info than the Github page.
https://fulcrumresearch.ai/2025/10/22/introducing-orchestra-...
The demo video in the GH page didn't work for me, but there is also one on Twitter/X: https://xcancel.com/fulcrumML/status/1984054489851310191
Submitted a PR with AWS Bedrock support: https://github.com/fulcrumresearch/quibbler/pull/5/files (credits!)
Is an anthropic api key really necessary? A major roadblock for taking a test drive. Already have a Claud Max subscription but an anthropic api key still need at least 5$/mon extra.
You could use the prompts in the code to create a Claude Code sub-agent[1], which would do the same thing without an API key
How would you invoke the subagent? Can a HookResponse cause a subagent to be invoked, to perform analysis on the action taken and then inject that back into the main loop?
Or would the hook invoke another instance of claude code?
I just read through the hook docs and I’m a bit fuzzy on the bidirectionality of it.
I really want Anthropic to let me make an API token that pulls from the same pool of usage that my Pro subscription does with the official clients. It would be cool to be able to run experiments with alternate clients and automation and stuff without having to go swipe the card at the ol' API token refilling station.
Sounds like video streaming services…
An agent's agent?
A gentleman's gentleman!
Sounds like a wip to me, "do it better or get punished"
Can the same concept be used for non coding tasks?
Cool concept, but I picture Quibbler as an off-brand Batman villain.
Replace the middle manager
hey HN! happy to answer any questions
this kind of tool is especially useful in longer running tasks to enforce your intent without having to check in on your agent all the time
Can users stack Quibblers, so Quibbler 2 corrects Quibbler 1 if, say, it fabricates an issue in the code it's reviewing? If so, have you found an optimum number of Quibblers for the Quibbler stack? Also, might users form a Quibbler council such that multiple Quibblers review the same thing and form a consensus before proceeding?
I love the pixel-perfect precision with which this comment is straddling the Poe's Law line.
That aside I also love the concept of Quibbler Council and I'd get a kick out of seeing it in action.
MoQs - Mixture of Quibblers? Would be convenient to have them run on dedicated FGPAs. Then they can facilitate near real-time quibbing at the network level across all packets.
But who polices the vibe police?
“Well, who’s gonna monitor the monitors of the monitors?” — Enemy of the State (Movie)
> We’ve found Quibbler useful in preventing agents from: 1) fabricating results without running commands
What a world we've created for ourselves
Next step is critics for the critics.
Until they reach critical mass.
Vibeception