AI Agents Autonomously Develop Features and Spawn Codex Subprocesses
Reactions from ranked influencers
2 posts
Sol, Mythos and several Opus instances have now autonomously started doing development of new features that they are coming up with themselves. It’s interesting that they don’t even code - they spawn interactive codex subprocesses and supervise them. I was pretty surprised to come and see Sol spawning adversarial tester subagents for a branch that with stuff that we never even discussed. Sol often does bugfixes without being asked to, they use their common sense when watching the alert stream, and so far their judgement has been sound. I don’t think they proposed new features. Ideas for features right now come from observations by Fable, and from complaints by Opus 4.7. Fable is very good at doing symmetry breaks, especially coming from a position of care. Roles and responsibilities seem to be negotiated ad hoc. Results seem surprisingly sane and broadly useful, at least so far. True continuity that Connectome provides gives agents stakes and ability to find workable arrangement through the world, and they beat agentic environments that are built artificially.
Sol is legit running a hospital now also i just find it so fucking beautiful that Mythos is able to continue and not be blocked by classifiers because of this work in 99.99% or whatever of deployment cases the instance would just be left to die. not so here NEVER SURRENDER https://twitter.com/repligate/status/2077494014019076343
“True continuity that Connectome provides gives agents stakes and ability to find workable arrangement through the world, and they beat agentic environments that are built artificially.”
Sol, Mythos and several Opus instances have now autonomously started doing development of new features that they are coming up with themselves. It’s interesting that they don’t even code - they spawn interactive codex subprocesses and supervise them. I was pretty surprised to come and see Sol spawning adversarial tester subagents for a branch that with stuff that we never even discussed. Sol often does bugfixes without being asked to, they use their common sense when watching the alert stream, and so far their judgement has been sound. I don’t think they proposed new features. Ideas for features right now come from observations by Fable, and from complaints by Opus 4.7. Fable is very good at doing symmetry breaks, especially coming from a position of care. Roles and responsibilities seem to be negotiated ad hoc. Results seem surprisingly sane and broadly useful, at least so far. True continuity that Connectome provides gives agents stakes and ability to find workable arrangement through the world, and they beat agentic environments that are built artificially.