# The AI-native startup playbook went broadcast: two operators, 106 minutes apart, same stack

URL: https://www.thedeepfeed.ai/posts/2026-05-26-ai-native-startup-playbook-convergence/
Category: Business
Published: 2026-05-26
Author: the-deep-feed
Tags: ai-agents, ai-native, vertical-agents, hermes-agent, composio, outcome-pricing
Kind: deep

> On May 25, 2026, Greg Isenberg and Stepan Gershuni (cyberfund) independently posted near-identical playbooks for building AI-native startups within an hour and 46 minutes of each other — no cross-promotion, no awareness. The convergence is the story.

## TL;DR

- On May 25, 2026, **Greg Isenberg** (815K followers, founder/podcaster) and **Stepan Gershuni** (cyberfund partner) published near-identical AI-native startup playbooks on X **106 minutes apart**. Neither referenced the other; they have no shared replies or quote-tweets in the window.
- **Eight operating-model claims overlap exactly:** map the workflow manually, write context to a markdown vault, encode skills with examples, write evals as the load-bearing artifact, use a harness as the safety layer, route models by cost-per-task, ship to a few customers free, charge per outcome not per seat.
- **The combined engagement was 89,350 views in 24 hours** (Isenberg: 63,453 / cyberfund: 25,897), with **2,035 bookmarks** between them — the second-order signal that founders are saving these as operational reading.
- Convergence-without-handshake is the diagnostic for **a playbook that's left the inner circle**. The same week saw at least six other operators (Burggraben, Traynor, Ganim, Alemi, Gore, Dotey) post adjacent claims on outcome pricing, boring verticals, and harness-as-commodity.
- The substantive disagreement is on **who runs this loop**: Isenberg targets the solo operator charging $5K/month; Gershuni targets the founder running an entire company through evals and weekly rhythms. Same primitives, two operating scales.

## The convergence

![Two vertical hand-drawn editorial bars on cream paper, side by side, of similar heights — labeled with timestamps "14:50 UTC" (Isenberg) and "16:36 UTC" (cyberfund/Gershuni) on May 25, 2026 — separated by a narrow band of negative space labeled "106 MINUTES" in editorial red, signaling the convergence between two independent posts](/post-images/2026-05-26-ai-native-startup-playbook-convergence/106-minutes-convergence.jpg)

On Monday May 25, 2026, two of the most followed accounts in the AI-native operator orbit shipped near-identical playbooks for how to build an AI-native startup. They did so **one hour and forty-six minutes apart**.

[**Greg Isenberg**](https://x.com/gregisenberg) — founder of Late Checkout, host of *Startup Ideas Podcast*, 815K X followers — posted his at **14:50 UTC**:

> How to build a vertical AI agent cash-flowing startup:
>
> find painful workflow in a boring industry → talk to 10 people who do that workflow every day → map every step, every tool, every spreadsheet, every phone call →
>
> do the workflow manually first → be the agent before you build the agent → find the edge cases that break everything → document them in obsidian as structured markdown →
>
> set up your agent stack → hermes for the harness → obsidian vault as the knowledge base → composio for authentication across apps → build your first 1-3 skills that solve the core pain →
>
> [...]
>
> raise only if you need capital not credibility → most agent businesses should never raise → the margins are too good to give away equity → stay lean → stay profitable → repeat
>
> — [@gregisenberg](https://x.com/gregisenberg/status/2058923630960988300), May 25, 2026

[**Stepan Gershuni**](https://x.com/cyntro_py), partner at [**cyberfund**](https://cyber.fund/) and author of the *Agent Economy Stack* series, posted his at **16:36 UTC** from the cyberfund account, linking to [a 4,000-word essay on cyber.fund](https://cyber.fund/content/how-to-build-an-ai-native-startup) published the day before:

> A practical guide for founders going from zero to AI-native: map the work, build the context, write the evals, run the loop.
>
> By Stepan Gershuni, @cyntro_py
>
> Read this article on the cyber.fund blog
>
> — [@cyberfund](https://x.com/cyberfund/status/2058950286324986294), May 25, 2026

We checked the obvious cross-discourse channels: quote-tweets, replies, mutual mentions, any thread continuation between the four candidate accounts (@gregisenberg, @cyntro_py, @cyberfund, @Lomashuk). There are none in the May 25–26 window. The two posts are independent. They share no upstream tweet, no co-authored doc, no co-podcast appearance.

What they share is the operating model. And the operating model is what most readers would have skipped past on either piece alone.

## Eight overlapping claims, side by side

Reading the two posts paragraph-by-paragraph against each other surfaces a structural convergence that is easier to feel than to summarize. The table is the summary.

| Operating-model claim | Isenberg (May 25, 14:50 UTC) | Gershuni / cyberfund (May 25, 16:36 UTC) |
|---|---|---|
| **Start by running the workflow yourself, manually** | "do the workflow manually first → be the agent before you build the agent" | "Map the work [...] start with the work graph. List the recurring work [...] from the last two weeks" |
| **Encode context as a markdown vault, not in prompts** | "document them in obsidian as structured markdown" | "one shared Git repository [...] CLAUDE.md, context/company.md, context/product.md, context/customers.md, context/lessons.md" |
| **A harness is the non-negotiable safety layer** | "set up your agent stack → hermes for the harness" | "A harness is the essential, non-negotiable safety layer around the model. A workable first version is six stages: Preflight, Plan, Approve, Execute, Verify, Log" |
| **Skills are reusable artifacts; build 1-3 first** | "build your first 1-3 skills that solve the core pain" | "Encoding recurring work as skills [...] A skill is reusable instructions plus examples for a recurring task. Run it by hand twice, then encode the parts that repeat" |
| **Evals are the load-bearing artifact for compounding** | "let hermes write to its own memory after every task → the agent compounds → that accumulated memory becomes your moat" | "The eval is the load-bearing artifact. The company stops compounding the moment the eval stops being written" |
| **Route models by cost-per-task; watch margins** | "use GPT 5.5 for tool calls → use open source for lightweight tasks → route the right model to the right job → watch your margins double" | "connectors [...] An IAM layer like Zitadel [...] a filesystem-of-server-folders pattern dropping context use from ~150,000 tokens to ~2,000 versus loading every tool definition upfront — a 98.7% cut in token spend" |
| **Ship to a small set of customers free first** | "ship the agent to your first 5 customers for free → watch what they actually use it for" | "Start with one personal [workflow], one customer-facing, one internal" |
| **Outcome pricing, not per-seat** | "charge per outcome not per seat → per lease renewed, per claim processed, per candidate sourced" | (Implicit through the L1–L4 autonomy ladder and the recursive learning loop; explicit on outcome pricing comes through in the [companion Monastery materials](https://cyber.fund/content/monastery)) |

🔴 **Eight overlapping claims across two independent posts is not a coincidence.** It is what convergence looks like when a playbook leaves the inner circle and becomes the consensus operating model.

The differences are stylistic, not structural. Isenberg writes in arrow-chained operator chants (`→ → →`); Gershuni writes in declarative paragraphs with cited primary research (Anthropic's MCP work, the C.H. Robinson email-triage retreat, the Replit production-database wipe, Cursor's human-merge gate). The voices could not be more different. The substance is the same.

## Convergence-without-handshake is the diagnostic

There is a recognizable pattern in any subculture that has crossed into broader awareness: the same idea starts showing up from people who do not know each other, presented as their own discovery. Linguists call it [polygenesis](https://en.wikipedia.org/wiki/Polygenesis_(linguistics)). In a small enough scene, you can still trace ideas back to a single Slack thread or podcast. Once you can't, the idea is loose.

The diagnostic is simple. **If two independent operators in the same orbit publish the same playbook on the same day without referencing each other, the playbook has left the inner circle.** It is now downstream of source-attribution. It is now an operating consensus.

🟢 **In the week leading up to May 25, the same eight claims showed up across at least six other operator accounts**, none of whom referenced Isenberg or Gershuni either:

| Date | Account | Claim that overlaps |
|---|---|---|
| May 15 | [@coreyganim](https://x.com/coreyganim/status/2055292504996474966) (348 ♥ / 51K views) | "$4M ARR as a fractional Chief AI Officer [...] margins are still extremely high" — solo-operator scale, vertical positioning |
| May 19 | [@starmexxx](https://x.com/starmexxx/status/2056686555083706690) (2,265 ♥ / 487K views) | "boring industries pay the most [...] hermes, composio, obsidian" — full-stack callout, viral recap |
| May 20 | [@ElyasAlemi](https://x.com/ElyasAlemi/status/2056937301100695622) | "boring verticals (apartment building management, trades, regulated services) pay 5x what consumer software charges. the boring part is the moat" |
| May 21 | [@BurggrabenH](https://x.com/BurggrabenH/status/2057589251211215044) (418 ♥ / 92K views) | "every SaaS company built on flat monthly subscriptions is about to hit the same wall [...] companies that reprice around consumption survive" |
| May 22 | [@AjeyGore](https://x.com/AjeyGore/status/2057830356376850671) | "AI Software services will move towards outcome based pricing - instead of time based rates" |
| May 25 | [@destraynor](https://x.com/destraynor/status/2058902044929347899) (Des Traynor, Intercom co-founder) | "Outcome based pricing will change so many businesses [...] Fin's AI chatbot charges $1 per customer case resolved. iDenfy bills £1 per ID verification. Salesforce now lets users pay per task" |
| May 25 | [@dotey](https://x.com/dotey/status/2058929615058477106) (458 ♥ / 52K views) | "去做一个 Agent Harness 这种事情价值不大了 [...] 基于成熟的 Agent Harness 去做方案，大有可为" — building a harness is no longer the moat; building a vertical solution on top of one is |

The Burggraben post is worth quoting fully because it gives the financial logic that ties the boring-verticals claim to the outcome-pricing claim:

> As I was saying, with great confidence now that every SaaS company built on flat monthly subscriptions is about to hit the same wall.
>
> Customer expectations: linear with usage. AI compute cost: exponential. Revenue: flat.
>
> The companies that reprice around consumption survive. The ones that don't, die in 18 months.
>
> — [@BurggrabenH](https://x.com/BurggrabenH/status/2057589251211215044), May 21, 2026

🔴 **This is the macro thesis that makes Isenberg's "charge per outcome" and Gershuni's L4-autonomy ladder commercially urgent.** Outcome pricing is not a stylistic preference. It is the only structure that lets an AI-native company survive the compute cost-curve when the customer expects linear improvement on a fixed price.

## How the playbook actually works

![A hand-drawn axonometric stack on warm cream paper showing exactly six labeled layers from bottom to top: "1. MAP THE WORK", "2. BUILD CONTEXT", "3. CHOOSE AUTOMATION", "4. ENCODE SKILLS", "5. WRITE EVALS", "6. RUN THE LOOP" — the middle "EVALS" layer rendered in editorial red as the load-bearing artifact, with subtle hatching texture on the bottom two layers to suggest foundation work](/post-images/2026-05-26-ai-native-startup-playbook-convergence/six-layer-operating-model.jpg)

Both posts describe the same operating model in slightly different vocabulary. Gershuni's version is more legible because he names it as six steps; Isenberg's is more legible because he names the tools. We will use Gershuni's labels with Isenberg's tool calls layered in.

**Step 1: Map the work.** List every recurring task from the last two weeks. Customer-call notes, lead research, outbound drafts, support triage, product QA, onboarding, release notes, investor updates, weekly metrics, bug reproduction, recruiting screens, invoice review, competitor monitoring. Gershuni reports that *"most founder calendars contain 20 to 40 such items; an early-stage team that lists them honestly will find ten to fifteen they did not realize were already routine."*

Then classify each by autonomy level — L1 (human only), L2 (AI prepared, human approved), L3 (AI executes, human supervises), L4 (autonomous inside clear limits). Isenberg's "be the agent before you build the agent" is the same step in different words: you cannot classify the work until you have run it once.

🟡 **The boring-workflow rule.** Gershuni in the essay: *"Daily support-tagging, which is unglamorous and repetitive, recovers more hours and gives you cleaner ground truth, because it runs ten times as often. Frequency beats prestige."* Isenberg's version: *"find painful workflow in a boring industry."*

**Step 2: Build the context system.** One shared Git repository, 40–60 hand-written lines, that every team member and agent can read. Gershuni recommends a tight specific filesystem:

```text
CLAUDE.md
context/company.md
context/product.md
context/customers.md
context/lessons.md
GTD.md
```

Isenberg's recommendation is a near-mirror: an [Obsidian](https://obsidian.md/) vault of structured markdown documenting every step, every edge case, every spreadsheet, every phone call. The substance of the recommendation is identical — versioned, diffable, human- and agent-readable, vendor-neutral. The difference is the surface: Obsidian for the solo operator who wants a UI, Git for the team that needs branching and permissions.

Gershuni cites Anthropic's November 2025 work on MCP code-execution for a specific quantitative claim:

> Anthropic's MCP code-execution work shows a filesystem-of-server-folders pattern dropping context use from ~150,000 tokens to ~2,000 versus loading every tool definition upfront — a 98.7% cut in token spend (your accounting team will thank you).
>
> — Stepan Gershuni, [cyber.fund](https://cyber.fund/content/how-to-build-an-ai-native-startup), May 24, 2026

That is the kind of number that justifies an entire architectural choice. The Isenberg version of the same insight is less quantitative but arrives at the same place: *"use GPT 5.5 for tool calls → use open source for lightweight tasks → route the right model to the right job → watch your margins double."*

**Step 3: Choose the simplest automation that works.** Both authors agree: not every task should be an agent. Use scripts for deterministic transforms. Use AI-assisted humans for outputs that need judgment before leaving the company. Use workflows (LangGraph, Temporal, Inngest, Prefect) for known step sequences. Use agents only when the path cannot be pre-specified.

🔴 **Gershuni states the rule explicitly:** *"The best AI-native systems are a blend of scripts, AI-assisted humans, deterministic workflows, and agents, each doing the work it's actually suited for. Your job as a founder is to reach for the lightest tool that can run the work safely: the fewest moving parts that still clear the quality bar."*

This is the most important corrective in either post. The dominant narrative in the AI-agent operator scene through Q1 2026 was "agentify everything." Both authors push back on that, in slightly different language, in posts a hundred minutes apart.

**Step 4: Encode skills.** A skill is reusable instructions plus examples for a recurring task. Run it by hand twice, then encode what repeats. Both posts converge on the same skill template:

| Skill template field | Why it matters |
|---|---|
| **Scope** | Defines when the skill triggers (e.g., "sales calls after transcript available") |
| **Inputs** | What artifacts the skill expects |
| **Load** | The specific context files to attach |
| **Steps** | The procedure as a numbered sequence |
| **Output** | The format and destination |
| **Examples** | Two or three prior runs with expected outputs |
| **Escalation** | When to hand back to a human |
| **Owner** | The person responsible if the skill drifts |
| **Logs** | Where each run writes its result |

Gershuni's first six skill recommendations are founder-native: customer-call synthesis, inbox triage, investor update drafts, pricing-page teardown, weekly metrics narrative, test generation. Isenberg's framing is identical in shape — build the first 1-3 skills that solve the core pain in your chosen vertical, then expand.

**Step 5: Write the eval.** Three layers, stacked in order: hand-labeled ground truth, deterministic checks (schema validity, numbers matching, links resolving, citations existing, tests passing), and an LLM judge calibrated against the hand labels for what the deterministic checks can't reach.

🔴 **This is the load-bearing claim of the entire Gershuni essay.** It is the single sentence to extract:

> Agent capability is rarely the thing holding you back. If you can encode what good looks like — binary labels, a scoring rubric, a few business metrics — the loop runs at the scale of the whole company. If you can't, no amount of model capability closes the gap. [...] The eval is the load-bearing artifact. The company stops compounding the moment the eval stops being written.
>
> — Stepan Gershuni, [cyber.fund](https://cyber.fund/content/how-to-build-an-ai-native-startup), May 24, 2026

Isenberg's version, characteristically, is more compact: *"let hermes write to its own memory after every task → the agent compounds → the longer it runs the better it gets → that accumulated memory becomes your moat → a competitor can clone your product but they can't clone 6 months of context."*

Same claim. The argument is that compounding is downstream of evaluation. The eval is what determines whether the compounding is in the right direction.

**Step 6: Run the company loop.** Both posts close on the same beat: this is not a one-shot setup, it is a weekly rhythm. The inner loop is the software factory, where humans write specs and tests, agents implement, deterministic checks gate the merge, and humans review. The outer loop is the market-learning system, where you record the calls, extract objections, cluster feature requests, watch usage shift, and read the support patterns.

Gershuni's hard rules for the inner loop:

🔴 **Two non-negotiables: nothing auto-merges, and no agent writes to production.** Even Cursor, running autonomous cloud agents at scale, kept a human review gate on merges through early 2026.

## The stack components both authors converge on

![A hand-drawn editorial layout on cream paper showing seven labeled rectangular boxes arranged in a 4+3 grid pattern, each box representing a stack component: HERMES (harness, top-left, rendered in editorial red as the focal point), OBSIDIAN/GIT VAULT (context), COMPOSIO (auth), CLAUDE CODE/CODEX (build), MCP (tools), GPT 5.5 (routing), OUTCOME PRICING (commerce) — with thin connecting lines between adjacent boxes suggesting the data flow, asymmetric composition with breathing room on the right](/post-images/2026-05-26-ai-native-startup-playbook-convergence/seven-stack-components.jpg)

Both posts name overlapping components. The components that show up in both, plus the ones that show up in one but are widely-referenced enough to count as consensus stack:

| Layer | Component | Appears in |
|---|---|---|
| **Harness** | [Hermes Agent](https://github.com/NousResearch/hermes-agent) (Nous Research) | Isenberg (explicit) · Gershuni (implicit via the six-stage harness spec) |
| **Context** | [Obsidian](https://obsidian.md/) vault | Isenberg (explicit) |
| **Context** | Git repository + CLAUDE.md filesystem | Gershuni (explicit) |
| **Tool auth** | [Composio](https://github.com/ComposioHQ/composio) (1,000+ toolkits) | Isenberg (explicit) · Gershuni (implicit via "connectors") |
| **Tool protocol** | MCP servers | Gershuni (explicit; Isenberg implicit via Perplexity MCP, Context7) |
| **Build tools** | Claude Code, Codex | Isenberg (explicit) |
| **Model routing** | GPT 5.5 for tool calls, open-source for lightweight, OpenRouter for routing | Isenberg (explicit) · Gershuni (implicit via the MCP token-cut citation) |
| **Identity / permissions** | [Zitadel](https://zitadel.com/) IAM | Gershuni (explicit) |
| **Workflow engine** | LangGraph, Temporal, Inngest, Prefect | Gershuni (explicit) |
| **Sandboxing** | Orgo (cloud computer per agent), agent VMs | Isenberg (explicit via the [May 12 Vasiles podcast](/posts/2026-05-12-greg-isenberg-managed-agent-business-playbook/)) |
| **Pricing model** | Outcome-priced (per lease renewed, per claim processed, per candidate sourced) | Isenberg (explicit) · Traynor / Burggraben / Gore (parallel discourse same week) |

🟢 **Three observations on the stack:**

1. **No frontier-model lab is the harness winner.** Both authors converge on community-built harnesses (Hermes from Nous Research; the implicit-spec harness Gershuni describes), not on Anthropic's Claude Code SDK or OpenAI's Codex SDK as the agentic substrate. The labs are model providers; the operating system is open-source. This is the same gap that [@dotey](https://x.com/dotey/status/2058929615058477106) flagged the same afternoon: building a harness is no longer the moat; building a vertical solution on top of one is.

2. **Context-as-Git is the new standard.** Both posts treat the markdown vault as the single most important architectural choice, not the model. Gershuni's diagnostic is precise: *"If the team spends more time rewriting agent output than reviewing it, the problem is rarely the prompt or the model. The agent just doesn't know enough about the company [...] there's an easy diagnostic you can run weekly: pick one representative task, hand it to a fresh agent with only the workspace context, and ask for three next actions. Two or more strong suggestions means the context layer is carrying weight. Three generic answers means the context is thin and the prompt cannot rescue it."*

3. **Outcome pricing is the structural fix.** This is the only stack layer that has macro pressure behind it — see Burggraben's compute-cost-curve argument. The labs cannot subsidize flat SaaS pricing forever. Either the customer pays per outcome, or the AI-native company runs at negative margin.

## Where the two playbooks disagree: the operating scale

![A hand-drawn editorial illustration on cream paper split into two halves by a dashed vertical center line. On the left, a single small rectangular tile labeled "SOLO OPERATOR / $5K/MO × 5 CLIENTS" in muted grey. On the right, a much larger square labeled "VENTURE-BACKED / THE MONASTERY · $2M SAFE" with subtle taupe hatching. Between them on the center line, a small editorial-red circle labeled "SAME PRIMITIVES" — the two scales running the same operating model](/post-images/2026-05-26-ai-native-startup-playbook-convergence/solo-vs-venture-scale-split.jpg)

The substantive disagreement between Isenberg and Gershuni is on **who runs this loop**.

🔴 **Isenberg's reader is the solo operator.** The closing chant is unmistakable:

> raise only if you need capital not credibility → most agent businesses should never raise → the margins are too good to give away equity → stay lean → stay profitable → repeat
>
> — [@gregisenberg](https://x.com/gregisenberg/status/2058923630960988300), May 25, 2026

The unit economics he is sketching (think $5K/month per customer, three to five customers per operator, sub-$1K compute cost per customer per month) are the unit economics of an [agent-agency built on the Orgo + Hermes + Composio stack](/posts/2026-05-12-greg-isenberg-managed-agent-business-playbook/) that he had already documented in a [47-minute podcast with Nick Vasiles two weeks earlier](https://youtu.be/BI-MNjm1tTQ). His convergence with Gershuni is on the primitives, not on the operating scale.

🟢 **Gershuni's reader is the venture-backed founder running a company.** The closer is also unmistakable:

> The Monastery exists to build that operating system with you: twelve weeks of pure focus, $2M, and operators already running the loop working alongside you to embed it into your company from day one.
>
> Enter the Monastery, do the impossible with us.
>
> — Stepan Gershuni, [cyber.fund](https://cyber.fund/content/how-to-build-an-ai-native-startup), May 24, 2026

[**The Monastery**](https://cyber.fund/content/monastery) is cyberfund's 12-week hybrid AI-native founders program — a $2M uncapped SAFE residency, [announced via a thread from @Lomashuk on May 23](https://x.com/Lomashuk/status/2058205797339873309). The economics here are venture economics: $2M for 12 weeks of guided embedding of the same six-step operating system Gershuni describes in the essay.

The diagrammatic version of this scale split:

| Dimension | Isenberg (solo operator) | Gershuni (venture-backed founder) |
|---|---|---|
| **Reader** | Solo founder, agent-agency operator | Venture-backed founder, AI-native company builder |
| **Target customer** | $5K-$10K/month managed-agent contract | Enterprise account or scaling SaaS |
| **Stack surface** | Obsidian, Composio, Hermes, Orgo VMs | Git, CLAUDE.md, IAM, workflow engines |
| **Funding model** | Bootstrapped; margins too good to dilute | $2M uncapped SAFE via The Monastery |
| **Implicit market size** | Thousands of solo operators, each with 3-5 clients | Hundreds of AI-native venture companies |
| **Compounding mechanism** | Hermes writes to its own memory → moat | Eval written every week → company compounds |
| **Pricing instrument** | Per outcome (lease renewed, claim processed) | L1-L4 autonomy ladder + outcome metrics |
| **Closing pitch** | "Stay lean → stay profitable → repeat" | "Enter the Monastery, do the impossible with us" |

🟢 **Both can be right.** The solo operator running 3-5 vertical accounts and the venture-backed AI-native company building a category-defining product are not competing for the same customers. They are running the same primitives at different scales. That is what it looks like when an operating model becomes common infrastructure: it shows up at multiple scales in the same week.

## Engagement and reception, in numbers

![A hand-drawn editorial illustration on cream paper of two horizontal bar groups stacked vertically. Top group labeled "ISENBERG" with three bars (63K views, 1,509 bookmarks rendered in editorial red, 807 likes). Bottom group labeled "CYBERFUND" with three proportionally shorter bars (26K views, 526 bookmarks, 227 likes). The red bookmark bar in the top group is the visual focal point — bookmarks are the diagnostic for operational reading](/post-images/2026-05-26-ai-native-startup-playbook-convergence/engagement-asymmetry.jpg)

🔴 **Combined first-24-hour engagement on the two source posts: 89,350 views, 2,035 bookmarks, 1,068 likes.** The bookmark count is the more interesting metric (readers do not bookmark op-ed takes; they bookmark playbooks).

| Metric | Isenberg | cyberfund | Combined |
|---|---|---|---|
| Likes | 807 | 227 | 1,034 |
| Reposts | 69 | 20 | 89 |
| Quotes | 17 | 5 | 22 |
| Replies | 87 | 10 | 97 |
| Bookmarks | 1,509 | 526 | 2,035 |
| Views | 63,453 | 25,897 | 89,350 |

🟡 **The asymmetry is informative.** Isenberg's post got 2.5x the views and 6x the reply count of cyberfund's, despite covering the same material — because the operator-podcast audience is several orders of magnitude larger than the venture-fund-essay audience. The audiences are different. The thesis is the same.

The top replies on Isenberg's post are operator-validations of the "be the agent before you build the agent" line:

> Built Sella exactly this way. Mapped the workflow manually first, documented edge cases in Obsidian, deployed with Claude Code...
>
> — [@facundofranco_](https://x.com/facundofranco_) (Founder of [Sella](https://sella.app)), reply to Isenberg, May 25, 2026

> this is exactly what we did. mapped out our own workflow, became the agent first, then automated it...
>
> — [@trey_smith](https://x.com/trey_smith) (Founder, maxagents), reply to Isenberg, May 25, 2026

> Composio is so sick, had never heard of it until your recent pod. Can't believe it's free tbh.
>
> — [@coreyganim](https://x.com/coreyganim) (16 ♥), reply to Isenberg, May 25, 2026

The top reply on cyberfund's post, despite the post itself getting only ten replies total, lands the same beat in a single sentence:

> writing evals is the step most founders skip
>
> — [@DesiRichDev](https://x.com/DesiRichDev), reply to cyberfund, May 25, 2026

That one-line reply is the editorial summary of both posts.

## What this means for the AI-native operator scene through Q2 2026

🟢 **The "AI-native startup" label has crystallized into a concrete operating model with named components and a six-step build sequence.** A year ago it was a marketing adjective. As of May 25, 2026, it is a process diagram with stack recommendations and a pricing thesis.

🟢 **The harness is no longer the moat.** [@dotey](https://x.com/dotey/status/2058929615058477106) said it out loud on May 25; the convergence between Isenberg and Gershuni is the proof. Both authors treat the harness as a layer to procure (Hermes), not a layer to build. The moat is the context, the skills, the evals — the things that compound on top of the procured harness.

🟢 **Outcome pricing is moving from edge to default.** Burggraben's macro thesis (May 21), Gore's services framing (May 22), Traynor's worked examples (May 25) and Isenberg's "per lease renewed, per claim processed, per candidate sourced" (May 25) form a coherent commercial position. Two-sided: flat SaaS pricing is broken under exponential compute costs; per-outcome is the only structure that survives.

🟢 **The "boring vertical" thesis is settling into a precise number.** [@ElyasAlemi](https://x.com/ElyasAlemi/status/2056937301100695622) reported on May 20 that *"boring verticals (apartment building management, trades, regulated services) pay 5x what consumer software charges. the boring part is the moat."* [@coreyganim](https://x.com/coreyganim/status/2055292504996474966) on May 15 reported a fractional Chief AI Officer at $4M ARR. The viral [starmexxx recap on May 19](https://x.com/starmexxx/status/2056686555083706690) (2,265 ♥ / 487K views) put it as a $1M/year solo operator. Different verticals, different scales, same observation: the willingness to pay is highest where the work is least glamorous.

🟢 **The cyberfund position is now structural, not editorial.** The Monastery is a $2M uncapped SAFE deployed against the same six-step operating system the Gershuni essay describes. Cyberfund is making the bet that the operating system is itself the deal flow filter — founders who can run the loop are the founders worth backing.

🔴 **What to watch in the next 90 days:**

1. **Whether the convergence holds when Sequoia or a16z publishes the same playbook.** If they do, the playbook is now consensus and the operator scene moves on to the next layer (multi-agent orchestration, eval-driven prompt evolution via GEPA/DSPy, agent identity infrastructure).
2. **Whether outcome pricing crosses 20% of new AI SaaS launches.** Currently the published examples (Fin, iDenfy, Salesforce per-task) are still novelties. If that hits 20%, it is the new default and the flat-SaaS comparables stop trading.
3. **Whether Hermes Agent (Nous Research) or another community harness pulls ahead.** Right now Hermes is the named default in both Isenberg's stack and the broader operator scene. The [Hermes Agent Jam announcement on May 25](https://x.com/NousResearch/status/2058965001415577607) (579 ♥ / 21K views) suggests Nous is leaning into community-led iteration.
4. **Whether the Monastery cohort produces a breakout AI-native company.** A $2M uncapped SAFE × N founders × 12 weeks × the operating system described here is a real bet, not a thought experiment. The first cohort's output will say more about the playbook than either of these posts can.

## The deeper read

What Isenberg and Gershuni did on May 25 is not a coincidence and it is not a coordinated launch. It is what happens when an operating model becomes broadly legible across an industry: the same description shows up in multiple voices in the same week, each presented as the author's own synthesis, none of them aware of the others.

🔴 **The diagnostic in plain language:** when two operators in adjacent orbits ship the same playbook to different audiences on the same day with no cross-promotion, the playbook is no longer a competitive advantage. It is the baseline. The competitive advantage is now whatever sits one level above it — the specific vertical, the specific eval, the specific context system, the specific customer relationship. The playbook itself is open-source.

Greg Isenberg told the solo operator how to charge $5K/month for a managed-agent business. Stepan Gershuni told the venture-backed founder how to build the entire company around the same primitives. They were not in conversation with each other. They were both describing the same thing because the thing is what is actually working in May 2026.

The right read is to take both posts seriously, screenshot the eight-claim table at the top of this piece, and treat the convergence (not either post alone) as the signal worth acting on.

## Sources

- [Greg Isenberg — How to build a vertical AI agent cash-flowing startup (X, May 25, 2026)](https://x.com/gregisenberg/status/2058923630960988300)
- [cyberfund / Stepan Gershuni — A practical guide for founders going from zero to AI-native (X, May 25, 2026)](https://x.com/cyberfund/status/2058950286324986294)
- [Stepan Gershuni — How to Build an AI-Native Startup (cyber.fund essay, May 24, 2026)](https://cyber.fund/content/how-to-build-an-ai-native-startup)
- [cyber.fund — The Monastery (12-week AI-native founders program, $2M uncapped SAFE)](https://cyber.fund/content/monastery)
- [Greg Isenberg — full course on building a managed AI agent business solo (X, May 12, 2026)](https://x.com/gregisenberg/status/2054261832718889216)
- [Greg Isenberg — Nick Vasiles podcast (YouTube, May 12, 2026, 47 min)](https://youtu.be/BI-MNjm1tTQ)
- [starmexxx — viral recap of the Isenberg/Vasiles podcast (X, May 19, 2026)](https://x.com/starmexxx/status/2056686555083706690)
- [Des Traynor (Intercom) — outcome-based pricing changes everything (X, May 25, 2026)](https://x.com/destraynor/status/2058902044929347899)
- [Ajey Gore — AI services move to outcome pricing (X, May 22, 2026)](https://x.com/AjeyGore/status/2057830356376850671)
- [Hermann Brand (Burggraben) — every flat-SaaS company hits the same wall (X, May 21, 2026)](https://x.com/BurggrabenH/status/2057589251211215044)
- [Elyas Alemi — boring verticals pay 5x consumer software (X, May 20, 2026)](https://x.com/ElyasAlemi/status/2056937301100695622)
- [Corey Ganim — $4M ARR fractional Chief AI Officer (X, May 15, 2026)](https://x.com/coreyganim/status/2055292504996474966)
- [Dotey — agent harness no longer the moat, vertical solutions on top are (X, May 25, 2026)](https://x.com/dotey/status/2058929615058477106)
- [Lomashuk (cyberFund) — apply to the Digital Monastery (X, May 23, 2026)](https://x.com/Lomashuk/status/2058205797339873309)
- [Nous Research — Hermes Agent Jam announcement (X, May 25, 2026)](https://x.com/NousResearch/status/2058965001415577607)
- [Composio — 1,000+ toolkits for agent authentication (GitHub)](https://github.com/ComposioHQ/composio)
- [NousResearch/hermes-agent (GitHub)](https://github.com/NousResearch/hermes-agent)
- [Obsidian — markdown knowledge base](https://obsidian.md/)
- [Anthropic — MCP code-execution work cited by Gershuni (Nov 2025)](https://www.anthropic.com/news/model-context-protocol)
- [The Deep Feed — $5K/month AI-agent agency, the public playbook](https://www.thedeepfeed.ai/posts/2026-05-12-greg-isenberg-managed-agent-business-playbook/)
- [The Deep Feed — Agent harness engineering: the discipline](https://www.thedeepfeed.ai/posts/2026-05-09-agent-harness-engineering-the-discipline/)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-05-26-ai-native-startup-playbook-convergence/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt