@mattzcarey we use a system atlas that maps out top level all our repos then ast treesitter service to transverse by functions finally rg contents of functions by relevance. taking only the context that’s required
@omarsar0 How do you keep context from drifting as agents churn through work in real time and update state but the graph doesn’t update at the same speed due to embeddings?
@joanathanlmc the lookalike vendor one is sneaky. we hit something similar building AP review agents, ended up getting more out of tightening the review criteria than swapping models. curious if you catch those with rules or let the model judge
in the process of coming up with materials for my param grinder bot, i made @grok facilitate a multi-round debate between Claude and Codex. to make messaging event-driven, I just had each agent do a local git commit, and the following agent would pass that as part of their prompt.
@KSimback I ask my grok bot to create a bookmark sorter, classifies everything, summarizes then looks for relevance in my codebases. Finally creates Linear tickets which I review then apply a label which triggers my deployed runners to spin up agents to build it and PR.
After spending over 3 hours on a call with @bot today triaging work and sending it to our software factory to build I’m convinced this is a big deal for knowledge work and beyond.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
https://t.co/r0qFLdvd6c