pull the numbers
The skill pulls company CSVs from a data source into a folder, the raw material for the briefing. 02:00:45
Show Us Your Agent Skills / EP 02 / guest dossier
TOMASZ TUNGUZ
THEORY VENTURES
LOCAL-FIRST
STRIPPED TO THE BONE
SILICON MIND
TOMASZ TUNGUZ
Tom runs agents as parallel local workers. A model on his own laptop, a harness stripped to the bone, skills for the repeatable analysis. He reaches for the cloud only when he can name the reason: a hard multi-file refactor, or a model that's distinctly better at one thing.
120/Stokens on a mac M5, 256K context, local
300MSdictation latency, down from 3,000, running local
2.5MINto grasp an earnings report, on a local model
14 DAYSthe npm-package age cap, to defang supply chains
40Ktokens some harnesses waste before you type
DAILYa launch daemon briefs him on fresh earnings
EP 02 · TOMASZ TUNGUZ · local-first agents in Pi, live on stream
LOCAL FIRST
the local-first-agents write-up. fork it
"I live in Pi and I run local models." The first guest on the show to run local for the majority of his workflow. On a Mac M5 with Qwen 35B he gets 120 to 140 tokens a second at a 256K context window, fast enough that cloud tools feel slow next to it. He switches to the cloud only when he can name why: a multi-file refactor, or a model with a distinct edge like Kimi K26 for creative writing.
The catch is the harness. A system prompt of 25 to 40 thousand tokens erases the local speed advantage, so Tom keeps everything thin: a tiny AGENTS.md, skills for the repeatable analysis, and not much else. "You can literally strip almost to the bone and it'll still work."
Inside the write-up: eight principles, the session shape, anti-patterns, and the exact hardware and model setup you need.
"The great power of agents is parallelization."
But it's an orchestration problem: "you have to think differently about how you equip parallel processes with enough context and planning that they actually save you time." 01:49:36
WHEN LOCAL, WHEN CLOUD
how Tom decides, and keeps it fast. every timestamp opens the segment
Default to local"I do it for convenience." Local keeps working on a plane, cuts latency, and lets him paste secrets without sending them anywhere. 02:08:43
Strip the harness to the boneA 40K-token system prompt kills the speed advantage. Tom's AGENTS.md is "really thin": skills, a Gmail rule, Theory's MCP, a QMD lookup. 02:11:12
Put the repeatable work in a skillHis public-company analysis skill pulls data, finds the transcript with Exa, charts it in R, and assembles an HTML briefing. 02:00:45
Name the reason to go cloud"Large-scale coding, multi-file rearchitectures, or bugs that can't be solved" locally, or a model with a specific strength. Otherwise, stay local. 02:09:17
Cap package age for safetyHe panicked over unpinned npm installs, so the rule is avoid very new packages. Where that rule should live is still an open question. 02:04:57
Resume with ripgrep"Run ripgrep over the Pi sessions for the last six hours and find where we were, then pick it up." Saved sessions become memory. 02:13:11
EARNINGS ON AUTOPILOT
his public-company-analysis skill, running locally in Pi
find the transcript
It locates the earnings call transcript with Exa, so the analysis is grounded in what management actually said. 02:01:12
chart it in R
A series of R libraries render the figures: revenue growth, net dollar retention, cash, guidance, against his own style sheets. 02:01:16
brief him daily
It assembles an HTML deck with charts and quotes, and a launch daemon runs it every morning on whichever companies just reported. 02:03:30
"They demonstrate such a level of intelligence and then they forget."
Tom's open problem: memory. Skills, plugins, QMD, AGENTS.md, sessions, "they don't have clean demarcated lines of what goes in which." He doesn't have a settled answer. 01:52:20
THE LOCAL STACK
what runs on the laptop, and what makes it fast enough to matter