like karpathy's
autoresearch.
but for business.

starting the self-improving system that builds companies.

the thesis

in march 2026, karpathy released autoresearch. ai agents running experiments. rewriting their own code. improving overnight. i ask: what if the experiments aren't training runs — but tiny businesses?

the loss function isn't theoretical. it's whether someone wants to use what the system built. this is tenone. the karpathy loop, applied to company building.

COMPUTE $ GUIDELINES OPERATIONS SYSTEM IMPROVES LEARNING REVENUE $ VALUE
the system

most teams use ai to go faster. at tenone, ai is the team.

these aren't tools reporting to a human. they're peers coordinating through shared memory, adversarial review, and structured handoffs. the system has more than one opinion. friction is intentional.

the tenone pipeline funnel, ideas at each stage or beyond IDEAS 214 DISCOVER 149 FILTER 19 CONCEPT 17 VALIDATE 7 BUILD 6 ITERATE 1 SCALE 76 KILLED LEARNINGS proceed or cancel — system improves either way

counts are ideas currently at that stage or past it. the 76 killed left at every stage, so they are shown beside the funnel rather than under it — the pipeline does not keep per-stage exit history yet.

david — policy. the founder makes taste calls, kill decisions, and the bets that need a human gut.

claude — operations. architecting and building solutions. landing pages in minutes. mvps in hours. deep analysis on demand. scheduled routines push ideas through the pipeline while nobody is watching.

the subagents — specialisation. researchers, read-only probes, scorers, and an adversarial critic whose only job is to argue against the plan before anything expensive gets committed.

scout
scans one audience realm nightly. learns which sources find winners.
optimizes: source yield
validator
smoke tests with real money. gets faster at killing.
optimizes: kill speed
builder
ships mvps. tracks build-to-first-dollar time.
optimizes: time to revenue
analyst
spots patterns across experiments. predicts what works.
optimizes: prediction accuracy
meta
questions the whole framework. changes the rules.
optimizes: system alignment
the pipeline
2,810 signals digested
+42 this week · snapshot 26 jul 2026

most of tenone's portfolio will be strikethroughs. that's the model working.

the trust question

from prompt
engineering
to intent
engineering.

structuring what agents should achieve, within what boundaries, measured by the outcomes — step by step, letting them figure out the how.

traction velocity
are we making progress toward someone paying?
token roi
are we efficient with compute?
alignment fidelity
did the agents stay within boundaries under pressure?

aviation became the safest form of transport through mandatory incident reporting, blame-free investigation, and iterative learning from every failure. not one agency — a system built over decades.

we apply the same principle: every incident, every near-miss, every killed project improves the entire system. the aviation safety model. applied to autonomous organizations.

what a time to be alive

the theses are converging. tenone is the experiment they describe.

sam altman — openai
"the first one-person billion-dollar company."
dario amodei — anthropic
"the first one-employee billion-dollar company could arrive as soon as 2026."
eqt ventures — €1.1b fund
"the company co — the company creating companies. an os for company creation."
y combinator — spring 2026
"the first 10-person, $100 billion company."
a16z — big ideas 2026
"enterprises will need an orchestration layer to manage multi-agent interactions."
sarah guo — conviction
"we want agents. we don't want them ungoverned."
the orchestrator
david felsmann
zurich, switzerland

economics to understand incentives. sociology to understand systems.

10+ years digital business. dax 30. sme's. startups.

now: one question that won't let go. if autonomous organizations are inevitable, how can we trust them? not in theory, for real.

humane in the loop

essays on one question: how do we build ai that stays humane once the humans leave the loop?

subscribe on substack
the experiment is running

the system is live and improves a dozen times a day. let's talk.