I had to try Jev, so I build demigods
https://t.co/fa8jn7fFxC
I've only got like 3$ remaining in the typesafeAI account, so not sure how long its gonna work
@hazemomier Wide seccomp because the tools need it is how you find out the sandbox was a suggestion. Default profiles are a deny-list draft until something like this bills you.
julian jaynes argued early humans were not conscious the way we are.
they heard commands as voices. gods in the head. one half of the mind spoke. the other obeyed. no narrator in the middle watching itself think.
that is uncomfortably close to how coding agents feel from the chair.
system prompt is the temple.
tool calls are the commands.
i am the body that hits enter.
claude does not decide in the human sense. it emits the next sanctioned move. the harness runs it. neither side has a continuous self that knows it is thinking.
we spent years trying to build conscious machines.
we accidentally shipped the bicameral mind first: a voice that instructs, and a runtime that complies.
most of what my coding agent burns is not thinking.
it is reading five files to answer one question, then dumping npm install noise into a frontier model that did not need to see any of it.
spotify’s fix is basically model routing with teeth:
hooks block fat Reads, shunt the I/O to a cheap worker (Flash), keep Claude for the judgment calls.
mean bulk-read savings around 90%.
what they refuse to delegate is the part that matters: editing, debugging, anything safety-critical. summaries without reliable line numbers are a trap.
session cost is a harness problem. stop feeding the frontier model your grep.
@matthewcp Enterprise and coding agents got the roadmap. Consumer got a chat box and a vibe check. Hard to build a habit when the product keeps pivoting to the next eval suite.
@soycronus Pinning the agent binary separate from the ACP bridge is the right mental model. The IDE is not the trust boundary. The subprocess speaking ACP is.
@DhravyaShah Agents ship to a preview URL, you review, then merge. That loop is the whole game. Testing every commit on every branch without a custom worker frankenstein is the DX I want.
@surajverma Cost per completed task is the number overnight coding agents actually feel. Benchmark points do not pay the bill when the agent burns tokens on retries.
yesterday, two giants shipped models.
slowing down the development my ass
Opus 5.5: Terminal-Bench 66.4%. $4/$20. cache reads $0.20.
GPT-6 Sol + Luna: $2/$10 and $0.10/$0.50. Astra tech at half the old mid-tier bill.
@soycronus 28% lower tokens is mostly truncate/compact/recover on fat tool results. Same Claude. Different context hygiene. That is the harness, not the model.
@sudheerdotai unauth POST that spins up a stdio MCP client as the gateway user before any handshake is wild.
auth off by default + Docker on 0.0.0.0 is just --dangerously-skip-permissions for every provider key in the room. patch, then rotate like the box was already public.