Garry Tan Spends $100K to $200K a…

1. Key Themes

Theme 1: Personal AI agents with deep memory are moving from chat tools to chiefs of staff

Agents differ from chatbots by holding vast context about you

Tan's agent has ingested decades of personal data, and YC runs an internal agent with institutional memory. The core argument is that working memory, not intelligence, is the human bottleneck agents remove.

"I think the number one thing that people get wrong is they just think that AI is just chat and then it is so much more than that."

"it's read all of my email and all of my Dropbox. It's read like all of my Apple notes like you know even my therapy notes going back 20 years."

"A person tracks seven plus or minus three things at once, he says, while an agent holds about 1,000 pages, roughly three Harry Potter books."

YC is rebuilding its own operations around an agent

Straylight gives YC a "company brain" over a high-volume intake funnel, a template for how organizations may restructure.

"YC runs an internal agent called Straylight, with a personality and a company brain that holds every founder YC has backed plus every application and email. That intake alone is 80,000 to 90,000 applications a year for about 800 spots, and Tan says YC is rebuilding how it operates around the agent."

Tan's end-state vision is a proactive "Jiminy Cricket"

The product vision is an agent that anticipates, not just answers.

"His picture of the finished product is Jiminy Cricket on your shoulder, an agent that knows you, runs ahead and asks 'what about this?'"

Theme 2: Token costs are falling fast, making elite-level AI usage mass-market within years

Today's frontier usage is a "privilege," but deflation is steep

Tan spends $100K-$200K/year, yet unit prices are collapsing and he expects broad adoption soon.

"I spend about maybe a hundred to two hundred thousand dollars a year on tokens. And so, I realize that's a crazy privilege that not everyone can do that."

"Opus 4 launched in May 2025 at $75 per million output tokens, and Opus 5.5 costs $20, a 73% drop in 16 months. At May 2025 prices, $150,000 of Opus work would have cost $562,500."

Task-level economics are already cheap

Individual high-value workflows cost almost nothing, so the gap is mostly spend discipline and setup.

"10 hour-long podcasts come to about 125,000 words, and reading them on Opus 5.5 costs about 90 cents, so a full prep run lands near $1."

"A lighter version with 6 everyday workflows costs about $170 a year at current Claude prices, by The AI Corner's estimate."

"it seems inevitable in about two to three years, I would say, everyone's going to do it and it's going to feel awesome."

Theme 3: A fight over openness will determine who owns the agent layer

Frontier labs are consolidating into the "new big tech"

Tan sees the labs as incumbents-in-the-making and voices concern about concentration.

"I think you can fairly say that they are the new big tech."

"Friends inside the labs talk about 'the five kings', a future where almost all commercial value collapses into 5 companies."

Open source and open markets are Tan's counterweight

He wants a "Linux for personal AI," frames the next browser war around agents, and argues the US is ceding open-source ground to China.

"what I want is there to be Linux, right, for personal AI."

"He argues the US is handing Chinese labs a big lead in open-source AI, a gap you can see in Kimi K3's Opus-level release."

Platform gatekeepers could block the agent layer

The Open App Markets Act is positioned as a startup-survival issue, with Apple's iMessage restrictions as the example.

"the next Apple can't emerge right if Google Apple and all the trillion dollar companies don't open a lane and create markets we will live in a world where there aren't startups."

Theme 4: AI's labor impact is a "raise the floor" opportunity for domain experts, not just a cost-cutting tool

Domain expertise plus agent-enabled builders creates new companies

Tan's answer to job loss is increased agency and new company formation, with GStack as the enabling tool.

"I want them to be able to start a company in that space. And maybe it's exactly what they did before. Maybe it's something two clicks away. But either way, it requires an increase in the amount of agency."

"pair a medical biller who knows the field inside out with a 22-year-old who ships software with agents, and they can start a company in that space."

Founders should start from unique lived experience

"what do you uniquely know? What life did you live? What obscure hobbies do you have? Like what have you seen that nobody else has seen?"

Tan warns against the lazy cost-cut response

"the responsibility for CEOs and people who have a lot of privilege is not to cost-cut. Like, that is the easy way."

2. Contrarian Perspectives

Coding-agent safety and repeatability can destroy the most valuable part of agents Conventional wisdom prizes predictability in agent tooling. Tan argues the opposite: over-optimizing for it strips out the human-like quality that makes agents powerful. The evidence cited is OpenClaw's traction, now surpassing React in GitHub stars.

"they're so worried about sort of repeatability and safety that they actually trade off all of the beautiful parts of working with an AI agent."

"OpenClaw now has about 394,000 GitHub stars, more than React."

Social media's fix is a personal cognitive filter, not better platform policy Rather than regulating feeds, Tan proposes demoting the apps to infrastructure and letting individual agents curate. The apps become "just an API," and curation is personalized and cheap (about $2.56/year for a daily digest by the newsletter's estimate).

"you do have a social network but the social network is mediated through your own cognitive filter."

The AI market structure is already "two to three winners," and openness is a policy choice Tan rejects the idea that a handful of winners is a hypothetical risk. He says "we're already there" and argues policy, funding for new talent, open markets, and open source are the counterweights. His bar for winning is practical rather than ideological:

"it just has to be easy enough to use and cheap enough to use that everyone chooses it."

3. Companies Identified

Y Combinator

Description: Startup accelerator led by Garry Tan
Why mentioned: Case study of an organization rebuilding operations around an internal AI agent (Straylight)
Quotes: "YC runs an internal agent called Straylight... That intake alone is 80,000 to 90,000 applications a year for about 800 spots, and Tan says YC is rebuilding how it operates around the agent."

OpenClaw

Description: Open-source personal agent project created by Peter Steinberger, now under a foundation
Why mentioned: Exemplar of the human-like, memory-rich agent and evidence of developer demand for open pieces
Quotes: "OpenClaw now has about 394,000 GitHub stars, more than React, and Steinberger joined OpenAI in February 2026 as the project moved to a foundation."

GStack

Description: Open-source skill pack for Claude Code built by Tan
Why mentioned: Tan's answer to AI job disruption; claimed 400x founder output
Quotes: "It has about 141,000 GitHub stars, and Tan says it gets founders 400 times the output they had before."

Anthropic (Claude, Claude Code, Opus 5.5)

Description: Frontier AI lab
Why mentioned: Underlies the cost math and the 1M-token context window enabling long-memory agents
Quotes: "Claude Opus 5.5 shipped on September 22 with a 1 million token context window... Tan expects up to 20 books within a year."

OpenAI / Codex

Description: Frontier AI lab
Why mentioned: Named alongside Anthropic as "new big tech"; Codex is criticized with Claude Code for prioritizing repeatability over agentic qualities; hired Steinberger
Quotes: "Isaacson pointed to newcomers like Anthropic and OpenAI doing well against Apple and Google. Tan agreed, then reclassified them as the new big tech."

Apple

Description: Incumbent platform company
Why mentioned: Example of gatekeeping that blocks AI apps (iMessage) and the platform lane startups depend on
Quotes: "AI apps can't touch your iMessages, and Apple says only Apple can keep you safe enough. Tan calls that a smoke screen for Apple's market power."

Meta

Description: Social media incumbent
Why mentioned: Controls the feed algorithm that personal agents could displace
Quotes: "personal agents are a very direct way to build it, since Meta controls the algorithm today."

Google

Description: Incumbent platform company
Why mentioned: Cited with Apple as a trillion-dollar gatekeeper needing to open lanes
Quotes: "the next Apple can't emerge right if Google Apple and all the trillion dollar companies don't open a lane."

Kimi (Chinese open-source lab)

Description: Chinese AI lab releasing open models
Why mentioned: Evidence that China leads in open-source AI
Quotes: "a gap you can see in Kimi K3's Opus-level release."

The AI Corner

Description: Newsletter publisher selling the Personal AI Agent Kit
Why mentioned: Source of the cost estimates and the $170/year build
Quotes: "a budget planner that runs all 6 for $170.46 a year."

4. People Identified

Garry Tan

Description: President/CEO of Y Combinator
Why mentioned: Primary subject, a power user whose agent usage previews broad adoption
Quotes: "I spend about maybe a hundred to two hundred thousand dollars a year on tokens."

Walter Isaacson

Description: Biographer and host of The Builders
Why mentioned: Interviewer whose pushback (agent vs. data retrieval) and framing ("open source operating system") sharpened Tan's arguments
Quotes: "Isaacson's gloss was 'basically, an open source operating system,' and Tan agreed."

Peter Steinberger

Description: Creator of OpenClaw, now at OpenAI
Why mentioned: Cited by Tan as the model for human-like, memory-rich agents
Quotes: "Steinberger joined OpenAI in February 2026 as the project moved to a foundation."

Sen. Marsha Blackburn

Description: US Senator
Why mentioned: Introduced the Open App Markets Act (S.2153) that Tan backs
Quotes: "The bill, S.2153, was introduced on June 24, 2025 by Sen. Marsha Blackburn with bipartisan cosponsors and still sits in the Senate Judiciary Committee."

Nelson Cowan

Description: Cognitive psychology researcher
Why mentioned: His 2001 review puts human working memory at about 4 items, strengthening the case for agent memory
Quotes: "Nelson Cowan's 2001 review puts the human number at about 4."

5. Operating Insights

Centralize your data before you buy tools. The agent's value comes from context. Start with a single readable corpus of your company's customer interactions.

"Start where YC did and put every customer email and application your company has handled in one place your agent can read."

Pilot with one high-stakes meeting and benchmark against your own prep. This is a cheap, low-risk test of the agent's value, costing roughly $1 per prep run.

"Pick 1 meeting this week, hand your agent every document behind it, and compare its prep with yours."

Audit platform dependency and decide your agent posture deliberately. Choose between predictable coding agents and memory-rich assistants, and count how many features rely on one closed model or one platform lane.

"Decide on purpose where your setup sits: a coding agent for predictable output, or an assistant with memory that lives in your chat apps."

"Count how many of your product's features depend on a single closed model."

6. Overlooked Insights

Prompt-level transcription errors can silently corrupt agent workflows. The newsletter found its own interview transcript mangled product names, a reminder that agents inherit upstream data-quality problems.

"The first locked prompt fixes a mistake we caught on this very interview: its transcript turned Claude Code into 'cloud code' and GStack into 'G stack'."

Prompt caching is a major cost lever for heavy users. A single technique can cut an all-day coding loop's cost by roughly 73%, which matters more than model-price deflation for operators running agents continuously.

"The caching math that takes an all-day coding loop from $51,100 a year to about $13,650."