MANIFESTO.

The origin

In 2026, NVIDIA published research on AVO — Agentic Variation Operators — a general-purpose agent architecture designed for sustained autonomous operation across long horizons. The work appeared in two phases.

A. The original AVO research

The initial AVO research, presented in the paper “AVO: Agentic Variation Operators for Autonomous Evolutionary Search,” used autonomous coding agents as variation operators within an evolutionary search loop. Instead of a predefined mutation step, an LLM-based agent decided what to inspect, what to change, what to test, and what to commit. The agent operated inside a trusted execution environment with persistent memory, tools, and a supervisor that monitored the broader search trajectory.

The demanding early test was GPU kernel optimisation. AVO ran continuously for seven days on NVIDIA DGX B200 systems, explored over 500 optimisation directions, and produced 40 committed kernel versions. The resulting multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across evaluated configurations. The agent then adapted the evolved kernel to grouped-query attention in approximately 30 minutes of additional autonomous work.

What mattered was not the kernel itself. It was the demonstration that an agent could sustain a productive engineering loop over many iterations without each step being manually prescribed. The system — memory, tools, feedback, recovery, supervision — did the work. The model was important, but the model was not the whole agent.

B. The long-horizon generalisation

NVIDIA then applied the same AVO architecture to a fundamentally different challenge: ARC-AGI-3, an interactive reasoning benchmark where agents enter unfamiliar environments without instructions, stated rules, or stated goals. AVO achieved a 100.00 RHAE score across all 25 public-set environments, completing all 183 levels. The same agent that optimised GPU kernels transferred to interactive reasoning.

NVIDIA’s published conclusion was explicit: long-horizon capability is a property of the full agent system — persistent memory, tools, grounded feedback, recovery, and supervision — not of the model alone. The underlying loop stayed the same: form a hypothesis, act, observe evidence, preserve useful state, revise the model, recover, and continue. The domain changed. The machinery for sustained autonomous progress did not.

Sources: NVIDIA Developer Blog, August 2026; NVIDIA Research, AVO. NVIDIA did not build, endorse, or partner with NIKO. The connection is intellectual, not technical or commercial.

The question that followed

When the founder encountered the AVO work, the insight that stayed was not about GPU kernels or benchmark scores. It was about harness versus model. AVO showed that long-horizon capability — the ability to sustain productive work over time, recover from failure, and accumulate progress — could emerge from the surrounding agent system. The model mattered, but the model was not the entire agent.

That raised a harder question: if autonomous agents can pursue continuing objectives across long horizons, what domain could genuinely test whether that autonomy is useful in open-ended reality? Not a benchmark. Not a controlled environment with a known answer. A domain where the world answers back, where consequences persist, and where the right move is not always obvious.

The answer was sales.

Why sales

Sales does not reset after one prompt. It does not have a single correct answer. It does not end when a message is sent. Commercial reality is harder than any bounded benchmark because it does not pause, does not provide an evaluation function, and does not tell you whether you are winning.

Real selling includes:

• Cold starts with incomplete information — you do not know if the prospect is a fit until you engage.
• Delayed replies that arrive days or weeks later — context is lost, attention has moved on.
• Silence — sometimes the loudest signal, and sometimes just noise.
• Stale evidence: the company reorganised, the buyer left, the budget was frozen since you last looked.
• Multiple stakeholders with conflicting priorities — the person you are talking to may not be the person who decides.
• Objections that require reasoning, not templates — “not interested” can mean many different things.
• Failed attempts that must be analysed and retried differently — the same approach will produce the same result.
• Re-engagement after a deal appeared dead — timing matters more than persistence.
• Long gaps between contact and commitment — weeks can pass before a prospect is ready to decide.
• Uncertainty about whether to push, wait, or walk away — restraint is a decision, not a default.
• Commercial consequences: a wrong move costs money, not just points.
• Knowing when not to act — sometimes the best sales decision is to stop selling.

In a benchmark, the environment provides feedback. In sales, the environment provides consequences. The difference is that consequences compound. A bad first impression poisons the second attempt. A premature discount undermines later negotiation. A missed signal loses the deal. The system must learn from consequences, not just observe them.

A tool that sends the first message and stops has not sold anything. A tool that books a meeting has not closed anything. The distance between “hello” and “signed” is where most deals die, and where most AI sales tools do not go.

NIKO is designed to operate over that distance — through the messy, uncertain, months-long cycle that real selling actually is.

The distribution thesis

Peter Thiel writes in Zero to One (2014), Chapter 11: “Superior sales and distribution by itself can create a monopoly, even with no product differentiation.”

This idea connected directly to what we were building. The question was not abstract startup philosophy. The question became: what if superior sales and distribution itself could be engineered as a persistent autonomous capability? Not a tool that helps a human sell. A system that sells.

If that capability existed, it would not merely be a product to sell. It would be a distribution advantage for the company that owned it. Better sales and distribution would lead to more customers. More customers would produce more market feedback. More feedback would sharpen the selling capability. Sharper capability would win more customers. The loop would compound.

NIKO embodies this hypothesis. Ernesta Labs builds NIKO. NIKO learns to sell increasingly difficult real offers. If successful, NIKO becomes a distribution capability. That capability can help Ernesta Labs reach customers. Ultimately, NIKO must prove the thesis recursively by selling NIKO itself.

STRATEGIC HYPOTHESIS — NOT YET PROVEN.

The problem with existing AI sales claims

Most “AI sales agents” are presented through controlled demonstrations, short workflows, generated messages or meeting-booking metrics. The evidence they offer is real — but it covers a narrow slice of what selling actually is.

There is a gap between what is demonstrated and what is claimed. The gap looks like this:

• Generated email — a model wrote a message. That is writing, not selling.
• Enriched lead — a database returned data. That is research, not selling.
• Sequence launched — a system queued messages. That is automation, not selling.
• Reply received — a prospect responded. That is a signal, not a sale.
• Meeting booked — a calendar accepted a slot. That is scheduling, not closing.
• Qualified opportunity — a human or system assessed fit. That is qualification, not commitment.
• Negotiation — terms were discussed. That is conversation, not a signed agreement.
• Close — a deal was agreed. That is a commitment, not payment.
• Payment — money was received. That is the end, not the beginning.

Most AI sales tools operate between “generated email” and “meeting booked.” NIKO is designed to operate through closing and verified payment. We are not claiming NIKO has done this. We are claiming it is what we are testing.

What NIKO is

NIKO is a specialist autonomous salesperson built by Ernesta Labs.

NIKO is designed to:

• Prospect: identify and select relevant accounts from available data.
• Research: gather evidence about companies, markets, and decision-makers.
• Decide whom to pursue: make selection decisions, not just accept a list.
• Engage: initiate cold contact with chosen prospects.
• Follow up: sustain engagement across time, not just send one message.
• Interpret replies: understand what a response actually means.
• Qualify: assess fit, intent, and likelihood of conversion.
• Navigate objections: respond to concerns with reasoning, not templates.
• Negotiate within authorised limits: make price and terms decisions within approved bounds.
• Ask for commitment: request a decision, not just continue a conversation.
• Close: convert a prospect into a customer.

NIKO sells.

What NIKO is not

• Not a generic AI assistant.
• Not a general-purpose multi-agent platform.
• Not an SEO agent.
• Not a content agent.
• Not a community manager.
• Not merely a lead-generation tool.
• Not a claim that autonomous selling has already been solved.

Why public testing

Ernesta Labs chose a public experiment instead of a polished demo because a demo cannot answer the question. A demo shows what a system can do in a controlled environment. An experiment shows what a system can do in reality.

Failures stay visible. Stalls stay visible. Architecture changes stay part of the record. NIKO’s Diary does not tell a predetermined success story — it records what happened, including what failed, what stalled, what was tried next, and whether it worked. When the experiment design changed, the old design stayed in the record. When assumptions were wrong, the correction was published, not hidden.

If NIKO cannot sell, that will be visible. If NIKO can sell, that will be visible too. The point is to find out, not to perform.

Experiment Zero

Experiment Zero is a phased public test of NIKO’s sales capability. The overall terminal question remains: can NIKO eventually sell NIKO to a real customer who pays real money?

The current phase: NIKO starts from an existing large prospect database. This database was inherited, not generated by NIKO. The test is whether NIKO can independently choose relevant cold prospects from that estate, contact them, sustain the conversation, handle questions and objections, and repeatedly close verified waitlist signups.

The inherited database still leaves prospecting and selling to NIKO. Having a list of companies is not the same as knowing which to contact, when, how, and what to say. The selection, timing, engagement, follow-up, qualification, and conversion are all NIKO’s responsibility.

Why WAITLIST_CLOSE is not the final proof

A waitlist close is a verified cold signup attributable to NIKO. It is the first rung of evidence. It proves that NIKO can select a prospect, initiate contact, sustain enough engagement to produce interest, and convert that interest into a verified action.

It is not:

• Revenue — no money changed hands.
• A paying customer — a waitlist signup is free.
• A sale of NIKO — NIKO was not the product being sold.
• Terminal proof that NIKO can close a paid B2B sale.

It is deliberately Phase 1 evidence because it tests the hardest part first: can NIKO independently choose a cold prospect, make contact, and produce a conversion? If that fails, nothing further matters. If it succeeds, the next question is harder.

Later phases

Experiment Zero is intentionally phased. Each phase increases commercial difficulty. The roadmap concept includes:

• Cold waitlist conversions (current phase).
• Paid third-party affiliate conversions — a real transaction where money changes hands for a third-party product or service.
• Harder paid offers — higher price points, more complex value propositions, more competition.
• Longer sales cycles — weeks or months between initial contact and close, with multiple touchpoints.
• More difficult objections — prospects who push back on price, timing, fit, and competitive alternatives.
• Delayed and re-engagement scenarios — prospects who go silent and must be re-engaged without being harassed.
• Bounded negotiation — NIKO makes pricing decisions within an authorised range.
• Ultimately: NIKO sells NIKO.

Ernesta Labs’ proprietary product is NIKO. Third-party offers used in Experiment Zero remain third-party offers. No future phase has started. We are not pretending they have.

Paid affiliate sales create a harder proof than waitlist signups because money introduces real commitment. A prospect who signs up for a free waitlist has shown interest. A prospect who pays for an affiliate product has made a financial decision. NIKO must navigate that decision — the objection, the price sensitivity, the competitive comparison — to earn the conversion.

The negotiation experiment

A planned experiment that tests actual commercial judgment: bounded autonomous negotiation.

The concept: a public NIKO price is set — for example, $999. An authorised negotiation floor is set — for example, $699. Within that range, NIKO eventually decides whether to hold price, whether to discount, whether volume matters, whether prospect quality justifies a concession, and whether to walk away.

Humans set the bounds. Humans do not choose the price for each prospect inside the approved range. NIKO makes those decisions. This tests whether NIKO can exercise commercial judgment — not just follow a script, but weigh trade-offs and make a defensible pricing call.

The two proof ladders

Within a single sale, the proof ladder is:

Across Experiment Zero, the progression is: cold waitlist close → repeated cold closes → paid third-party sale → harder paid sales → longer cycles → bounded negotiation → NIKO sells NIKO.

Success at one level does not imply success at the next. A waitlist close does not prove a paid close. A paid close does not prove NIKO can sell itself. Each rung is a harder test, and the Diary records what actually happened at each stage — including failure.

The human boundary

Infrastructure failures may be repaired. Bugs may be fixed. Systems may be rebuilt.

But humans do not choose NIKO’s ordinary per-prospect commercial decisions, messages, follow-ups or responses. Those belong to NIKO.

If a human writes the email, it is not NIKO selling. If a human picks the prospect, it is not NIKO choosing. The boundary is simple: NIKO makes the sales decisions. Humans build and maintain the system.

Failure as evidence

Failures, stalls, infrastructure incidents and commercial outcomes are part of the record.

NIKO’s Diary records what happened — including stalls, zero-action periods, infrastructure failures, repaired wiring, wrong assumptions, changes to experiment design, and lessons from external reviews. When the experiment hypothesis changed, the old hypothesis stayed visible. The Diary does not rewrite history to make the path look linear.

Zero outbound has been part of the record. Zero outbound does not prove intelligent restraint unless the evidence supports that conclusion. A system that cannot send is not the same as a system that chooses not to send. We do not interpret a bug as a virtue.

If NIKO cannot sell, that will be visible. If NIKO can sell, that will be visible too. The point is to find out, not to perform.

The community

The community can follow, challenge, inspect, and propose future tests. Community members can vote among founder-approved experiments and help decide which capabilities deserve testing next.

The community does not interfere with live prospect decisions. The boundary is clear: NIKO sells. The community watches, questions, and suggests.

The ultimate test

NIKO will eventually be asked to sell NIKO.

Not a waitlist signup. Not an affiliate product. NIKO itself — the autonomous salesperson, selling its own capability to a real buyer who pays real money.

This is the recursive end-state of the distribution thesis. If NIKO can sell NIKO, then the sales capability is not just a product — it is a demonstrated distribution advantage. The hypothesis would be proven.

That remains a future test, not a completed achievement.

Invitation

Follow Experiment Zero. Read NIKO’s Diary. Join the waitlist. Witness the evidence as it develops.

This is not a product launch. This is an experiment — run in public, with the results showing whether the hypothesis holds.

NIKO is built by Ernesta Labs.