Garry Tan

Confidence 0.95 · 9 sources · last confirmed 2026-10-03

Garry Tan is President & CEO of Y Combinator (since 2023). Engineer by background — Stanford Computer Systems Engineering; Palantir employee #10; Posterous co-founder (sold to Twitter); built the first version of YC’s Bookface (internal social platform + portfolio knowledge base). The wiki’s canonical AI-founder-type-archetype operationalised (per Hu 2026) — runs 10-15 parallel Claude Code sessions, ~400 PRs in review, ships open-source toolkits at venture pace.

Promoted from Dangling to an entity page on 12 May 2026 after three substantive source mentions:

1. GStack 2026 — first-party demo of his open-source Claude-Code-toolkit.
2. ex-brain 2026 — names GBrain as a parallel concept to Karpathy’s LLM Wiki.
3. Liu 2026 — deep first-party-class architectural treatment of Tan’s GBrain (24 skills, 21 cron jobs, 17,888-page brain repo).
4. Stanford CS153 2026 — Tan’s Stanford guest lecture (20 May 2026) with YC GP Diana Hu. The bumped-and-promoted contribution is the agentic-primitives-map-to-company-structure framing (skills as employees, resolvers as the org chart, skillify as the 10-step compliance protocol, check-resolvable as audit, trigger evals as performance reviews); the 5-day Posterous rebuild on a $200/month Claude Max plan anchor (10 people / $4M / 2 years collapsed to one founder / 5 days / $200); the cross-modal evals as a Skillify built-in in-progress feature (Opus 4.6 + GPT-5.5 + Deepseek V4 cross-evaluating); the typed-knowledge-graph + in-progress epistemology layer GBrain extension.

Role in the wiki

Tan is the wiki’s first first-party-operating AI-founder-type-archetype anchor. Where Hu 2026 prescribes the AI-founder-type (still builds / still coaches / leads by example), Tan is the worked example. Three angles surface across the ingested sources:

1. As the operator of a packaged harness — GStack (Tan 2026)

GStack — open-source toolkit (github.com/garrytan/gstack) wrapping Claude Code with named skills:

• Office Hours: 16-YC-partners adversarial-review distilled at 10% strength. “Pressure-tests your idea before you write a line of code.”
• Plan / CEO Review / Adversarial Review: multi-step auto-fix of design-doc issues; tracked scored improvement.
• Design Shotgun: multiple AI design variants in ~60s as choices.
• Code Review: staff-level bug-catching after plan.
• SLQA / SL browse: Playwright+Chromium wrapped as a CLI (replacing “the worst piece of software I’ve ever used” — Claude-in-Chrome MCP).
• Ship: pre-merge gate.

Velocity: built in 3 weeks; “now has more GitHub stars than Ruby on Rails.” The wiki’s first founder-CEO-of-major-accelerator open-source-harness instantiation.

2. As the architect of an autonomous-knowledge-base — GBrain (Liu 2026 + OceanBase 2026)

GBrain — Tan’s second open-source project; 24 autonomous skills + 21 cron jobs + 17,888-page brain repo on Postgres + pgvector. Built for personal AI agents (OpenClaw, Hermes, Claude Code). README: “Your AI agent is smart but forgetful. GBrain gives it a brain.”

Architectural principles Liu names:

• “Thin harness, fat skills” — harness is ~200 lines (model execution + read/write files + safety); all intelligence in skill files.
• CLAUDE.md + RESOLVER.md routing — RESOLVER.md dispatches user intent to skills across six categories (always-on / brain ops / content ingestion / thinking / operational / setup). “Skill descriptions themselves function as the resolver. The model reads the descriptions and matches intent automatically. No explicit routing code needed.”
• Fat skill = workflow contract — each skill is “not a prompt template, but an entire workflow: when to fire, what to check, how to chain with other skills, what quality bar to enforce.”
• Always-on signal-detector — fires on every inbound message; “An unlinked mention is a broken brain.”
• Cron skills that run themselves — “each job is deliberately thin — the prompt is literally ‘Read skills/{name}/SKILL.md and run it.‘” Jobs respect quiet hours (11 PM–8 AM by default); enforce idempotency.
• Deterministic split — separates latent work (reading, synthesis, pattern recognition) from deterministic work (database writes, calculations, reproducible outputs). “Mixing them is how agents hallucinate.”
• Tan’s latest iteration (Liu cites): “Fewer fatter skills makes the resolver shorter, which itself is less context bloat. Short resolvers are better than long ones.” Trend: fewer, more comprehensive skills with branching parameters > many narrow ones.

GBrain is the wiki’s operationalisation of the LLM Wiki pattern augmented with autonomous-action skills — the “fat skills” / act corner of Liu 2026’s three-architecture decision framework.

3. As the AI-founder-type archetype operationalised at YC-president scale

Tan’s reported workflow (per GStack 2026):

• 10 to 15 parallel Claude Code sessions all at the same time.
• “About 400 PRs to review right now.”
• “10, 15, 20, sometimes 50 PRs in any given day, depending on the number of meetings I have.”
• “I have multiple open-source projects with tens of thousands of stars.”

Direct operationalisation of Hu 2026’s AI founder type archetype: “still builds, still coaches and leads by example. If you’re the founder, this needs to be you at the forefront.” Tan IS the worked example.

Notable framings Tan has contributed

• “ADHD CEO vs autistic CTO” model allocation — “Claude Opus 4.6 is sort of ADHD CEO. He’s the guy you want to get a beer with and he’s got a billion ideas, but when the going gets tough, you got to call in your autistic CTO and that’s Codex.” First-party-CEO articulation of model-allocation-by-personality-fit-to-task as a working practice. Extended in Tan & Hu / CS153 2026 to a cross-modal evals workflow: Opus 4.6, GPT-5.5, and Deepseek V4 evaluate each other’s inputs and outputs and feed structured ratings back to the original sub-agent for iterative improvement — “you can metaprompt to get something that is 10 times better than the first version.”
• “Level 7 software factory” (invoking Steve Yegge’s Gas Town eight-stages-of-dev-evolution-to-AI) — “There’s this idea of trying to get to a level 8 software factory and GStack does not get you to level 8, but I do think it gets you to level seven.” Two-source-on-Yegge convergence with Böckeler 2026.
• “Office hours as 10% strength YC partner review” — codified 16 YC partners’ adversarial-review approach as a single Claude Code skill.
• “Fewer fatter skills” — Tan’s latest iteration: “Short resolvers are better than long ones.”
• Agentic primitives ARE company structure (CS153 2026) — “a skill is a squishy human being who’s an employee who has a capability. A resolver is the org chart — who handles what, how does it happen, the filing rules. Where it goes in the brain is the internal process. Check-resolvable is audit and compliance. A trigger eval is performance reviews.” The wiki’s first founder-CEO-altitude rhetorical claim that the agent-harness primitives are not merely metaphorically like a company — they are structurally identical to a company’s employees / org chart / audit-and-compliance / performance-review systems.
• 5-day Posterous rebuild on $200/month (CS153 2026) — “I was able to create like everything all the software we made over two years with 10 people and all that capital, but me with a $200 a month cloud code max plan and anyone in this room could do that and it didn’t take like two years, it took about 5 days.” The single most concrete 1000x-engineer anchor the wiki holds — operationalises Tan’s earlier “10× to 100× to 1000× as productive” citation of Steve Yegge into a specific founder-self-reported time and cost collapse.
• Boil the ocean (CS153 2026) — “my response to that is, actually, let’s boil the ocean — you can do the work of about 500 to a thousand people. And if that’s true, then all of the expectations that we currently have in society around what a founder can do, what a company can do, what a small team can do… they’re actually a thousandx wrong.” In productive tension with Onshore 2026’s wedge-first counter-position (“the idea of trying to boil the ocean all at once is very challenging — it has been an incredible benefit for our business to be great at one thing really early”); the wiki carries both vantages.

Career timeline

• Stanford — Computer Systems Engineering.
• Palantir — employee #10; engineer / designer / product manager.
• Posterous — co-founder; micro-blogging platform; sold to Twitter.
• YC — built first version of Bookface; later promoted to President & CEO.
• 2023–present — YC President & CEO.

Notable projects

• GStack — open-source Claude Code harness toolkit; github.com/garrytan/gstack.
• GBrain — open-source autonomous-knowledge-base; 24 skills + 21 cron jobs + 17,888-page brain repo.
• Bookface — YC’s internal social platform + portfolio knowledge base (first version by Tan).
• Conductor — multi-session Claude Code orchestrator Tan demos GStack inside of (third-party tool; Tan is the worked-example user).

Convergence with wiki sources

Source	Connection
GStack 2026	First-party demo; the wiki’s primary Tan source
Y Combinator 2026	The AI founder type archetype Hu prescribes is operationalised in Tan’s reported workflow. Same accelerator, same architectural prescriptions, two consecutive days of publication (Tan 23 April + Hu 24 April)
ex-brain 2026	Names GBrain as a parallel concept to Karpathy’s LLM Wiki. Two-source first acknowledgment of GBrain
Liu 2026 (AI Advances)	Deep first-party-class architectural treatment of GBrain — thin harness / fat skills / RESOLVER.md routing / signal-detector / cron architecture
Karpathy 2026	GBrain extends Karpathy’s LLM Wiki pattern with the act layer; structurally complementary architectures
Böckeler 2026 (Thoughtworks)	Two-source Yegge Gas Town convergence: Böckeler invokes Yegge as audience-orienting; Tan invokes Yegge with the specific level 7 vs level 8 terminology

Open questions

• GStack and GBrain GitHub repos — both are open-source; primary-source ingest of RESOLVER.md, THIN_HARNESS_FAT_SKILLS.md, skill schemas would substantiate the thin-harness/fat-skills framing.
• GStack adoption metrics — “more GitHub stars than Ruby on Rails” is a velocity claim from April 2026; 3-month / 6-month adoption arc worth tracking.
• GBrain enterprise extensions — Liu notes GBrain “optimizes for one power user’s workflows, not organizational deployment.” If Tan or the YC community open-sources an enterprise extension, the pattern’s scalability becomes empirically testable.
• Conductor as a distinct entity — Tan uses Conductor as the multi-session shell for GStack. First-party documentation on Conductor as a separate tool is an open ingest target.
• OpenClaw and Hermes — named by Liu as the personal AI agents GBrain is built for. Both currently Dangling first-mention; primary-source targets.

The Startup School keynote: the doctrine stated in full (Aug 2026)

The wiki’s fullest first-party statement of Tan’s position, and the source that converts the scattered GStack/GBrain material on this page into a single argument. Three things it settles:

• “Personal AGI” is the name he gives the thesis. Not an event but a diffusion — “AGI isn’t arriving as an event. It’s arriving diffused as your agent running on your context doing your work.” The equation is explicit: a rented frontier model + owned context + a harness that wires them = an agent that acts like a fast version of you. “Model quality is rented but your brain is owned.”
• GBrain’s scale is stated for the first time: ~220,000 markdown pages, 25 years diarized, “compiled mostly by agents, curated by agents, searched for by agents.” GStack is described as “mostly skill files plus a browser that the agents can drive” — the fat skills, thin harness slogan already in this page’s tags, now with the architecture behind it. See llm-wiki.
• The ownership politics, which is new. Skill files are “a piece of your cognition… extracted from your head, written down, and executable”, so the same 40 files are two different futures depending on whose repo they live in. “Own your skills, because if you don’t, your job becomes a skill file.” See ai-deskilling.

Two contributions are portable beyond his own stack. The latent-versus-deterministic diagnostic — “the most important question to ask here is where is the computation happening… confusing them causes every agent failure I’ve ever seen” — is now recorded on agent-harness as a failure taxonomy. And the skillify discipline (“if you have to ask for something twice, you failed”) is on agentic-engineering.

He is also unusually candid about the failure mode of his own architecture: “a brain nobody curates is a garbage dump with great search.” The stated remedies — provenance on every fact, contradiction checks, a librarian whose job is pruning — are the same three disciplines this wiki’s schema encodes, reached independently.

The investor-facing version, six days later (a16z, Aug 2026)

Tan restated the Startup School doctrine on a16z’s podcast six days after the keynote, and the pair should be read together — but the interview adds three things the keynote did not:

• Token maxing. “It costs, I don’t know, 50 or $100,000 a year to use the agents at full strength… you can live in 2028 today.” The frontier as a spending decision rather than a capability date.
• The maintenance tax on skillify. Asked whether errors compound across hundreds of markdown files, he names provenance and conflict resolution plus a scheduled sweep as the necessary machinery — see llm-wiki.
• “The harness wars” as his call for 2027, gated on inference cost: “it’s really a cost thing, it’s just too expensive right now.” Four weeks before YC’s own Paper Club made harnesses a research night.

And the heterodox note worth keeping, because it cuts against the rest of his own message: asked whether the constraints are intelligence-bound, Tan says “it’s all human” and predicts twenty years, not two — “there are real moats… a Microsoft isn’t going anywhere” — which he flags himself as “a little antithetical to what most people in Silicon Valley would say.” That sits in tension with Collison on the same YC stage a fortnight earlier.

The “20x company” framing (added 2026-09-28)

The New Way To Build A Startup (YC, Feb 2026): Tan names startups that automate all internal functions “20x companies” (a term from GigaML’s founders), presents them as an evolution of Parker Conrad’s “compound startup”, and lists three patterns: AI teammate, one source of truth, and a custom agent per employee.

Opening YC’s defense event (added 2026-10-03)

At The Startup Industrial Base in Washington (30 September 2026), Tan’s opening remarks carry his argument beyond software: “software has left the screen. It can see, it can navigate, it can fly, it can run entire factories”, and most of it, built for commercial reasons, “can be pointed at defense.” YC came to Washington, he says, because “we followed the founders”, and he sets the day’s question: “What would it take for a 20 person company to become a dependable supplier to the American military?” Later he interviews Transportation Secretary Sean Duffy and raises the safety of autonomy, citing a study in which self-driving cars reduce injuries by as much as 90%. Wardstone’s founder names Tan as the company’s YC partner. The event has no agent or harness content; it is the only source on this page about hardware and government buyers.