Inside YC's AI Playbook

Predictions and Expectations:

• Pete Kuman expects that in about two years, spending $100,000 or $1 million annually on tokens will become "commonplace" and cost only "a couple hundred bucks."
• Gary identifies that the "two-sentence description skill" will eventually become "better than i am" at writing them due to self-improving loops.
• Pete predicts that "most software in the future" will look like "just-in time software" that is extremely simple to start with and then extended by an agent.
• Pete believes that within "18 to 24 months" (potentially up to five years), a critical choice will emerge between centralized control where users cannot run their own prompts, or a decentralized model where users control their own software and prompts.
• Pete expects that in the future, "commercial software will come with this capability" of customizable, agent-extendable interfaces "out of the box."
• Pete believes that AI will "eliminate kind of the drudgery style work" and function as an "extension of yourself" rather than replacing humans.
• Pete fears a future where "personal computers never existed and there were only mainframes," where five "kings" control compute and data centers, preventing users from running their own code or changing prompts.
• Pete expects that a "one-time time warp" exists now that allows companies to "leapfrog every incumbent" and "all Fortune 500s" by adopting this open AI-native model.
• Pete believes that by defaulting to public broadcast of agent conversations, a "social control" will effectively keep private information private within high-trust environments.

Timelines and Milestones:

• YC started building their own harness and agents internally "about a year ago."
• Pete notes that the current "single player era of agents" (where harnesses are designed for a single human on a single machine) is ongoing, with the "multiplayer harness" for teams not yet "solved well yet."
• Pete predicts a shift in the token economics landscape "in two years," where costs drop from $100,000-$1 million to a "couple hundred bucks."
• Pete warns that the next two to five years represent a critical decision point for whether AI becomes centralized or decentralized.

Technology and Product Direction:

• YC plans to continue "record[ing] all the artifacts" (such as meeting recordings) to build a "shared organizational brain" and "connect our brains."
• The plan is to use AI as the "building layer for everything" rather than just a "co-pilot," moving away from "deterministic software wrapping in AI" toward "agent wrapping software deterministic tools."
• Pete intends to continue using "chat" as the primary interface because it is "the closest thing to human language" and "expression of thinking."
• Pete expects that "g-brain" and "hermes agent" will enable the ability to "run your own software," "change your own prompts," and "choose which model to use."
• Pete anticipates that future software will rely on "self-extending and self-referential" architectures similar to the open-source harness "pi."
• The strategy involves normalizing data into a "big table" format optimized for agent retrieval, utilizing "RAG," "graph RAG," and "hybrid RRF" with re-ranking.
• YC plans to expand the "tool registry" to include over 350 tools, allowing agents to manage tasks like "office hours," "booking journal entries," and "managing events."
• The direction includes implementing "autonomous self-improving loops" where agents review transcripts to improve skills, such as the "two-sentence pitch" skill.
• Pete believes that "multimodal" input (text, voice, pictures, files) is essential, noting that "chat interfaces" are sufficient for users to trust the engine to do more work without needing complex UIs.

Market and Industry Outlook:

• Pete observes that "agentic coding" is "really catching hold" with the introduction of tools like "Claude Code," "Windsurf," and "Cursor."
• The speaker notes that while many people still live in a world where data is siloed and complex questions take "several hours" to answer, the industry is moving toward "denormalization" where context is centralized for agents.
• Pete expects that companies that do not adopt these open AI-native practices will be "trapped" and "leapfrogged" by those that do.
• The outlook suggests a future where "every single person" can build these capabilities at their own company, raising the "floor" for new employees who can "automatically get a lot of the context" without waiting six months to ramp up.
• Pete predicts that the "rate of progression of ai" will be driven by "social cultural things" like the default assumption of recording meetings, rather than just technical factors.
• The speaker believes that the "multiplayer harness" problem is the next major frontier to solve, enabling "superpowers" at an "organizational level."

Company Plans:

• YC plans to maintain a "high trust environment" where "every agent conversation" is "globally viewable by any full-time employee" to facilitate transparency and learning.
• The company intends to invest "$10 to 100,000 a year on tokens" to fund these open, internal agent operations.
• Pete plans to continue building "just-in time software" by leveraging small code bases (e.g., "2,000 lines have marked down") and dynamic skills rather than massive monolithic codebases.
• YC plans to use the "skillify" meta-skill and "check resolvable" resolver to ensure skills are "D.R.Y." (Don't Repeat Yourself) and "M.E.C.E." (Mutually Exclusive, Collectively Exhaustive).
• The company plans to continue "denormalizing" data into a format optimized for "OpenClaw" or "Hermes agent" to allow agents to "ask any or answer arbitrary questions."
• Pete plans to keep the "tool registry" as a central primitive, allowing teams to add tools easily and making them available to both internal agents and individual agents on personal machines.
• The plan includes using transcripts from meetings to "write them back into the internal DB" and "internal CRM" to improve skills and context.

Financial Guidance:

• No specific revenue, earnings, or valuation guidance is provided; the only financial reference is the estimated operating cost of "$10 to 100,000 a year on tokens" for current AI-native operations.

Risks and Caveats:

• Pete acknowledges that the default state of most organizations is "command and control" and not open, which poses a barrier to this transformation.
• There is a risk that AI becomes "centralizing" with "five kings" controlling compute and data, preventing users from running their own software.
• The speaker notes that "security and privacy" concerns were initially a "blocker," requiring a choice to trust by default and be "willing to spend" on tokens.
• Pete warns that if companies "keep all the contacts locked down" because it feels "unsafe," they risk being left behind by organizations that are open.
• The speaker expresses concern that "two-sentence description" and other skills could fail if the organization does not have the "egalitarian" and "high trust" traits required.
• Pete notes that the "social etiquette" around recording meetings was previously a blocker, though it is becoming "default assumed."

Confidence and Disagreement:

• Pete is "really surprised" that non-technical finance staff could use the initial tools to ask real questions.
• Pete states with certainty that "chat is actually pretty good" and that he has "definitely changed [his] mind" about chat not being the UI for all AI applications.
• Pete believes with strong conviction that "this is the closest thing to us being able to connect our brains" and that the "two-sentence pitch" is a "needle pinprick" in the fabric of organizational operations.
• Pete is "not sure" about the exact timeline, suggesting it could take "18 to 24 months" or "five years" for the centralized vs. decentralized choice to fully manifest.
• Pete believes that the current state of "agentic coding" is "unbelievably powerful" and that worrying less about security (as he did with the database access) reveals this power.
• Pete expresses confidence that "OpenClaw," "Hermes," and "G-Brain" represent the "true personal AI moment" and the "Apple I moment" for the industry.
• Pete is "embarrassed to say" the M.E.C.E. term is from McKinsey but asserts the concept is "actually like the optimal resolver."