ExpertsFMHow To Build Superintelligence Inside Your Company
Y Combinator Startup Podcast · May 27, 2026
5 Selected Lessons:
framework · 3 heuristic · 2
Denormalize data into one big table/context store optimized for agent retrieval
Source
Y Combinator Startup Podcast — How To Build Superintelligence Inside Your Company (May 27, 2026) · at 6:45 ▸ · at 9:57 ▸
When this applies
when any AI system needs fast, flexible retrieval across data that currently lives in many normalized, siloed systems
The takeaway
Reshape your data architecture around the consumer (the agent), not around normalization theory built for humans and joins.
Why it works
Having all important context in one schema (vs many joined tables or third-party SaaS silos) lets an agent reason over everything at once without multi-hop retrieval failures.
One level up
Optimize infrastructure for the actual consumer of the system, even if it breaks prior architectural orthodoxy.
Where it stops holding
This stops working when data freshness/consistency requirements make a single denormalized store impractical, or when the dataset is too large/heterogeneous for one agent-optimized schema.
Detail
procedure:
- Consolidate important business data into one schema rather than many third-party SaaS silos
- Give the agent schema documentation
- Let it run read-only queries or retrieval against this single store
- For personal/other systems, denormalize into a schema optimized for agent retrieval (RAG, graph RAG, hybrid RRF, re-ranking)
rationale: It just turns out when all of that context is in one place, with a little bit of additional information about how the schema is laid out, an agent can go and answer arbitrary questions about the business.
In their words
It seems like you have to denormalize and do the big table thing, but specifically for the agent.
Common context layer + centralized tool registry for 'multiplayer' organizational agents
Source
Y Combinator Startup Podcast — How To Build Superintelligence Inside Your Company (May 27, 2026) · at 12:12 ▸
When this applies
when trying to scale a powerful individual tool into something usable by an entire team or organization
The takeaway
Infrastructure that is useful to one person becomes exponentially more valuable when centralized and made available to everyone's agents.
Why it works
A shared data warehouse plus a growing, de-duplicated tool registry lets any agent — whether a dedicated internal one or an individual's coding agent — access the same organizational capabilities, compounding value as more tools are added.
One level up
Solve for the 'multiplayer' version of any single-player superpower tool to unlock organization-wide leverage.
Where it stops holding
This stops working when teams build duplicate, siloed tools instead of contributing to a shared registry, recreating the original fragmentation problem.
Detail
procedure:
- Build a common context layer (data warehouse) with as much internal context as possible
- Build an internal tool registry that any agent can call
- Let every team add tools whenever they find a piece of work an agent could improve
- Expose the registry both to dedicated internal agents and to individuals' own coding agents
benchmarks: {"tools_at_start":20,"tools_now":"350+"}
rationale: The tool registry is what turns these agents into something that's useful at work.
In their words
The tool registry is where most of the like YC specific stuff lives, right? The tool registry is what turns these agents into something that's useful at work.
Skillify + resolver + MECE/DRY check for self-expanding agent capability
Source
Y Combinator Startup Podcast — How To Build Superintelligence Inside Your Company (May 27, 2026) · at 15:27 ▸
When this applies
when any ad-hoc solution proves useful and should be converted into a standardized, discoverable, non-duplicative process
The takeaway
Continuously convert validated improvisation into structured, de-duplicated reusable capability.
Why it works
Converting a one-off successful action into a named skill linked to a central resolver makes it discoverable and callable later; checking it against existing skills for MECE/DRY prevents capability sprawl and redundancy.
One level up
Systems that compound capability need both a mechanism to crystallize new learning and a mechanism to prevent redundant proliferation of that learning.
Where it stops holding
This stops working when no one periodically runs the dedup/MECE check, leading to overlapping skills that confuse rather than help; or when skills are crystallized before they're genuinely validated.
Detail
procedure:
- Do something novel in the agent session
- If it worked well, say 'skillify it' to turn it into a tool/method call
- Run a 'check resolvable' pass comparing the new skill against all existing skills/tools
- Verify it is DRY (don't repeat yourself) and MECE (mutually exclusive, collectively exhaustive)
- Consolidate into one parameterized tool rather than many overlapping ones
rationale: It's bad to have 10 skills that do all the same thing. It's good to have one skill or one tool that has parameters that then let you call them.
In their words
if you have a dry and MECE resolver table anywhere, it's actually like the optimal resolver.
Unrestricted access unlocks agent power more than narrow permissioning
Source
Y Combinator Startup Podcast — How To Build Superintelligence Inside Your Company (May 27, 2026) · at 4:38 ▸ · at 6:45 ▸
When this applies
when deciding how much scope/access to grant any new autonomous tool or employee
The takeaway
Excessive safety-driven restriction often costs more in lost capability than it saves in prevented harm, especially in high-trust environments.
Why it works
Narrowly scoped tools constrain what an agent can usefully do; broad access lets it synthesize across all available context to answer arbitrary, previously-unaskable questions.
One level up
Worry less about worst-case misuse and more about enabling best-case capability when the environment is high-trust.
Where it stops holding
This stops working when the environment is low-trust, externally facing, or when the cost of a single misuse event is catastrophic rather than recoverable.
Detail
rule: When tool scope frustrates an agent's usefulness, consider granting much broader access rather than incrementally expanding narrow permissions.
threshold: Tried narrow tools first, found them insufficient, then jumped to full database access 'late at night' as an experiment.
rationale: It turns out that the thing that was hampering the world was being worried about security and privacy and all the things that could go wrong; worrying less reveals unbelievable power.
In their words
So I was like, what if we just gave the thing like access, complete access to the production database where it could just like trample on anything?
Chat is the right default interface because it mirrors human thought expression
Source
Y Combinator Startup Podcast — How To Build Superintelligence Inside Your Company (May 27, 2026) · at 32:19 ▸
When this applies
when designing the interaction layer for any system meant to capture nuanced human intent flexibly
The takeaway
Favor the most expressive, least constraining interface for tasks whose scope is open-ended and trust in the underlying system is high.
Why it works
Because chat/language is the closest medium to how humans express thought, and because trust in the agent reduces the need to visually verify every step, chat scales to arbitrary tasks without constraining them into a fixed box.
One level up
Interface design should match the bandwidth of the underlying intelligence, not force it into pre-AI interaction metaphors.
Where it stops holding
This stops working when trust in the agent is low (so users need to visually verify/control each step) or when a task genuinely benefits from a fixed, purpose-built view generated just-in-time.
Detail
rule: Default to chat as the interface; generate specific UI/software just-in-time only when needed.
rationale: Chat is the closest thing to human language, and human language and writing is basically the closest thing to expression of thinking. So, chat is the closest stepping stone to clear intelligence.
In their words
Chat is actually pretty good because, like, you trust the agent, you increasingly trust the agent to do more the work, and you trust its decisions.