An agent powered by unstructured content in a vibe database is a loose cannon. It doesn't matter how amazing the models get.
We just launched Sanity Context. Your agent, mainlining your structured content in real time.
We're not pivoting to AI. We were always here for it.
https://t.co/aZ5xUcdglN
by recursively distilling events in this manner, the agent appear to stay coherent and directed over thousands of messages. in one sense, the quality of the context window is refined over time, rather than deteriorating.
@kmelve It is very early days, for sure, so I miss the maturity. It has some hiccups where it needs to be kicked a bit, the screen editor is buggy at times. And I am integrating it into MIRIAD right now and it remains to be proven that it can keep going for hours without the meltdowns.
Just discovered Letta Code. Most coding agent suffer complete cognitive collapse after prolonged sessions. This is a big limitation when running orchestrated agents as they flame out in a matter of minutes. Letta is actively curating its context window. No more "compaction"!
Our slack-like agent orchestration tool https://t.co/hJCFSGHJOy is proving fluid and flexible. Here I have aligned on a spec, then assembled a team of agents implementing it in about 25 minutes of autonomous labor. Collaborating agents with lean context windows are scary smart.
Why can't my coding agents—backend, frontend, architect, designer—just hash it out in a thread like people do?
We made a thing over the holidays to try it. MIRIAD is basically Slack for agents.
We are prepping for an open test run: https://t.co/hJCFSGHJOy