GIỚI THIỆU VỀ TẬP NÀY
Watch the video version for the demos and screen shares: https://www.youtube.com/live/FPhLBODA2LY
Brian Douglas and John McBride, co-founders of Paper Compute, joined Melanie and co-host Shy Ruparel to talk about what happens to agent work after the session ends. Brian led developer advocacy at GitHub and founded an open source company the Linux Foundation acquired, which is where he met John. John spent years in cloud infrastructure at Cloud Foundry, VMware and AWS and is Paper Compute's CTO.
Brian showed paperd, the daemon that records every agent session on the machine it runs on, and an unreleased feature for chatting with a teammate's session while it is still running. John explained tapes, the open source layer between the coding harness and the inference provider that captures everything crossing the wire. From there: skills as the new runbooks, whether skills survive as models improve, and two of Brian's side projects that make the case for small models, an agent that beat Pokemon Red and a Tetris benchmark where turning reasoning off won.
What we covered:
- paperd keeps sessions past the 30 days Claude Code holds them locally. The tagline: turn your dead tokens into skills.
- Chatting with a teammate's live session instead of asking for a screenshot in Slack. Single player to multiplayer.
- High fidelity traces: tapes captures what the harness never instruments, like Claude Code's yes/no permission check.
- A skill built from someone else's session is worth more than one from your own. Skills as runbooks: make a PR, resize cloud resources, spin up staging.
- Benchmarking teammates from their code review sessions, and a twin CLI that reviews a PR in a teammate's style.
- The Friday problem: one or two people per team turn sessions into skills, the other 25 never do. Knowledge transfer beat cost savings as the real unlock.
- John's hot take: skills evaporate into the frontier models. What lasts is usage insight plus evals, and prompt magic pushed down into deterministic code.
- Pokemon Red: state read from the ROM in a sandboxed container, no internet, no phoning a bigger model, a summary carried forward every thousand turns. It never talked to NPCs, then it beat the game.
- Every agent session is a heist: a crew of small local models routed by vLLM Semantic Router, about 31 cents of electricity a day.
- Tetris: frontier models paused to reason, small open-weight models with thinking off played better.
- Gateways as a waterfall of policies, and why cost is "just a label" unless you own your inference.
- New closing segment, does it pass the vibe check? The Odyssey, Groundhog Day, Jurassic Park and the Death Star.
Resources:
- Event page (Luma): LUMA_LINK_PENDING
- Paper Compute: https://papercompute.com
- tapes: https://github.com/papercomputeco/tapes
- Paper Compute Discord: https://discord.com/invite/T6Y4XkmmV5
- Brian's blog (every agent session is a heist): https://b.dougie.dev
- Brian's Pokemon Red dataset and adapters: https://huggingface.co/bdougie
- vLLM Semantic Router: https://github.com/vllm-project/semantic-router
