@trq212 genuinely curious how the rewrite came together. was Mythos part of it, or did publicly available models do the whole thing? both would be impressive in different ways, just want to know which.
@AvitalBalwit reminds me of Schopenhauer...art briefly turns us into "the pure, will-less subject of knowledge," free from the striving that usually defines us.
@trq212 managed agents bug instead of claude code, but getting a lot of spurious "Permission to use [mcp tool] has been denied." errors that makes it difficult to reliably do work in these sessions.
@ericzakariasson in an early glass build, cmd+t inside of terminal used to quickly spawn new terminal tabs that were easy to switch between. now, it opens a dropdown, and switching between terminal tabs in agent view is not as ergonomic.
convergence in the current regime is overdetermined: labs are months apart on a shared stack with a shared talent pool and shared compute vendors. that’s what you’d see whether or not singleton dynamics kick in later. the singleton thesis is ‘whoever crosses the threshold first pulls away faster than others can catch up.’ a tight race into the threshold is fully consistent with a runaway after it, maybe even predicts it. you’d need evidence that lead times are shrinking under RSI, not just that they’re currently small
imo nobody is seriously arguing that generating form is the same as solving a problem. the better question is how thinking gets distributed between designer and tool, and what understanding the designer retains.
there is a real cost in how cognitive labor gets divided between person and AI. but it doesn’t follow that AI outputs aren’t downstream of a thinking process. they are. the question is whose.
@jess__yan@michael_chomsky@AnthropicAI any chance u guys are planning to support something like programmatic tool calling or buffering outputs to a file when large for the managed agents MCP connector? this is the biggest blocker preventing me for replacing our custom Vercel Sandbox + Claude Agent SDK integration.
@mc_anthropic @trq212 super excited about the new managed agents release. i'm integrating with it rn and have two quick questions:
1. getting an internal service error whenever i upload files for the agent to interact with (`sesn_011CZtbT6wcMTaouy8Mt7XoZ` and `sevt_01E6vj3NRN9iH8esMQHzMZgA` for a sample session error)
2. it doesn't appear to use programmatic tool calling with the mcp tool calls, or the claude code harness technique automatically piping large mcp outputs to a file for programmatic access. this makes it more difficult to perform tasks like doing analysis and generating exports on large outputs from mcp tool calls.
any feedback or guidance would be greatly appreciated 🙏