Sam Hogan, founder and CEO of Inference.net, walks through his daily workflow, internal tooling, and how his 10-person team uses AI agents to automate go-to-market, engineering, and hiring tasks — spending roughly $20,000/month on model APIs to effectively add five “AI employees” alongside the human team.
Role shift: from technical founder to terminal-first CEO
Sam still writes code but focuses on automating non-technical work (go-to-market, recruiting, contract negotiation) using internal AI agents.
He treats the terminal as his primary interface because the team has wired GTM tooling, CRM, email, and document review into CLI-accessible workflows.
The bar for what LLMs can automate in traditionally non-technical work is moving fast, so he invests engineering time into internal tooling rather than waiting for off-the-shelf products.
Morning routine and personal productivity system
Wakes up, checks Slack on phone (team often active at 1–2 AM), then triages email for high-priority items — “the normal old-fashioned way.”
Uses a single Apple Note as his master to-do, daily standup, and reminder system; cross-platform, persistent for over a year, replaces Notion/Linear/Evernote/AI tools.
Consumes content via aggressively curated Twitter feed (~300 muted words), no AI summarization tools (“kitschy”), Twitter blocked 9 AM–5 PM and in bed (usually 10 PM–6:30 AM).
Gator: internal AI agent for email, CRM, and GTM
Built in a weekend during Opus’s moment (December); pure vibe-coded, plugged into Slack, email, CRM, DocuSign, and internal docs.
Handles email drafting, lead enrichment, contract review, and other GTM workflows; has a frontend dashboard for visibility.
Team plans to open-source parts of it; taught them how to build actually useful agents vs. demos.
Bern-E: multiplayer coding agent built on OpenCode
Spins up a Cloudflare sandbox, loads all repos with configurable skills; chosen over Codex/Claude Code because OpenCode is uniquely configurable and supports OpenAI models (Claude Code locks to Anthropic).
Multiplayer: any team member can join a session, prompt Bernie, or answer its questions in real time — rare among current harnesses.
Tradeoff: bleeding-edge customization vs. maintaining rough edges (low-level Cloudflare/library bugs); would switch to a good out-of-the-box tool when one exists.
Planning process: whiteboards → Linear → agent first passes
Small team (10 people): collaborative whiteboarding with co-founders, then team feedback; “if you have a good plan, the coding part is actually pretty easy now.”
Slack bot can ingest planning conversation transcripts, break them into tasks, and push to Linear.
Tickets can be assigned directly to Bernie, which spins up a sandbox and produces a first-pass PR — human reviews and finishes.
Where the human stays in the loop
Sam jumps between product/engineering threads and GTM/hiring threads; usually has at least one active project in each.
Spends most of the day in the terminal (Ghostty) because GTM tooling, contract negotiation, and lead review are all CLI-accessible.
Ghostty chosen for intentional design, ergonomic DX, theming/configurability, and cohesive modern terminal philosophy.
Hiring and recruiting with AI
Heavy Twitter sourcing: follows RL/cursive language model researchers, conversations lead to hires.
Gator dashboard includes hiring tab (job descriptions hosted on Ashby); AI ranks applicants by resume scoring + deep-research Google search across multiple axes.
“Most useful thing we’ve done with hiring”; every role (even SDR) expected to use AI — one sales rep built a Chrome extension for prospecting.
Thesis: just as PCs went from data-processing niche to universal knowledge-work tool, AI will become universal across all functions.
Using AI as editor, not writer
Drafts thoroughly in Apple Notes, uses AI for copy-editing/first draft, then moves to Google Docs to strip “AI voice” and make it sound human.
Writing matters for customer-facing content; AI-generated blog posts without human thought are avoided.
Sunk-cost trap with agents: starting a 20-min task with an agent, hitting edge cases at 40 min, continuing because “already close” — now more selective.
Why the team relies on agents less (last 2–3 weeks)
Conspiratorial but experience-backed view: model quality has dipped (labs deny weight-swapping/quantization, but periodic dips are real and observable).
Early heavy usage mapped the edge of capability; now more concise about what agents handle vs. what stays human.
Evaluation method: give agents a good plan + self-judgment criteria, let run 5–6 hours, inspect output to find failure modes.
Not everything is delegable yet; complex products at scale still need human steering.
Catalyst: self-service fine-tuning and deployment for small language models
Product launched same day as recording; opposite end of spectrum from giant generalist models (ChatGPT, Claude).
Small models: super fast, super affordable, do only a few things (e.g., specific summarization, data extraction/transformation) — don’t know capital of France.
At scale, pulling web/db/unstructured data through giant models is prohibitively expensive; Catalyst makes training/deploying specialized small models easy.
Internal dogfooding: Schematron (open-sourced Oct 2024) extracts structured JSON from HTML (e.g., “all person info from this page”); used in sub-agent scraping workflows; adopted by large companies.
R2 CLI: custom downloader for Cloudflare R2 (S3-compatible) because existing CLI worked but he wanted his own.
Drum sequencer: fun side project.
Figma clone: let model run overnight; works surprisingly well — draws shapes, handles stroke thickness, core Figma interactions functional.
Live bug bash on Catalyst launch day
Bug: starter project creation should seed training + evaluation datasets, but table shows empty.
Debug flow: check frontend logs (no 400s, PostHog error only) → suspect backend → write reproduction steps in Slack → tag Linear bot to auto-create ticket.
Co-founder Abe jumped on it immediately; Linear integration confirmed fix in progress.
Gopher (Linear bot) has “Gen Z” personality — swears, sassy; HR flagging is also automated.
10 humans, 5 bots: the ratio and the math
10 human team members, ~5 persistent bots (Gator, Bernie, Gopher, plus others) = ~30% AI “headcount.”
$20K/month on model APIs (February) ≈ cost of one extra person for five agent-equivalents.
Skeptical of “fully agent-run company” demos: complex products at scale need constrained agents with context; too many degrees of freedom leads to wild goose chases.
Agents are super useful but kept narrowly scoped; human-in-the-loop remains essential.
