Seven of this week's fourteen launches sell to the agent, not the person

Seven of this week's fourteen launches sell to the agent, not the person

Software has always been sold to the person who would use it. The seat was the unit, the login was the proof, and the pricing page counted heads. Two ordinary days of launches suggest that the person is quietly being replaced in that sentence — not as the buyer, but as the thing the product is built around.

Fourteen business launches cleared our bar on Product Hunt's organic daily rankings for 16 and 17 September 2026. Seven of them are sold to somebody who is building or running AI agents, and one of those seven exists so that an agent can buy from you. Two more take their instructions from inside an AI assistant. That leaves five that are ordinary business software in the shape the last decade would recognise.

That is a small sample from a single platform, and we say what it does not show at the end. But the pattern is worth naming, because it changes what a buyer is actually purchasing.

What launched, and what we left out

Both days carried a full ranked top seventeen with scores. Fourteen cleared our bar for business software, each checked against its own site before inclusion.

16 September: Weave Router, which routes each coding-agent request to the cheapest model that can handle it; Twigg, a hosted context store for applications that call language models; ZeroClick, which turns an API into a storefront an AI agent can buy from; Jottoo, meeting notes that become tracked tasks; Project Feed, project management with file review built in; and Convo, a sales copilot that works during the call.

17 September: Creem, merchant-of-record payments for digital products; Bitrise Remote Dev Environments, cloud Macs and Linux machines for developers and for coding agents; NovaSynth by Noveum, which sends synthetic callers at a voice agent; MCPJam, a testing and evaluation platform for Model Context Protocol servers; Figo, competitor monitoring with a Monday briefing; Pitchfire for Startups, which matches a pitch deck to venture firms and sends the introduction; AskDeck, which returns a PowerPoint and a narrated video from a spoken brief; and Higgsfield API, one asynchronous interface to more than fifty image and video models.

What we left out matters, because a selection is an argument. We excluded personal utilities and novelty: a cat photo app, an AI journal, two menu-bar apps, a Mac cleaner, a phrase vault, a minimal browser, a video recorder, a spatial board and a social feed of other people's coding-agent usage. We excluded a virtual-office product tagged under Games. We excluded a directory of agents that live in a personal phone contact list. We excluded a platform extension of an existing product, and a Mac notch widget that reads business metrics but is still one person's laptop.

We also excluded two things that were not excluded on quality. Appwrite 2.0 ranked second on 16 September and is a serious open-source platform, but Appwrite's own blog dates 2.0's availability for self-hosting to 7 September and 2.1 to 14 September, so the Product Hunt post is not the release. And an established observability platform launched a page that announced no dated release of anything, so there was nothing to date a launch to. Neither is a judgement about the product.

The seven that sell to somebody else's agent

Take them in order of how far from the person they sit.

Weave Router sits between a coding agent and the model providers. It scores how hard each turn is and routes it to the cheapest model that can get it right, across the subscriptions a team already holds, and only switches mid-conversation when the saving beats the cost of rebuilding the prompt cache. The customer is a team whose agents are already running; the product's job is the bill.

Twigg holds the conversation. Instead of an application rebuilding and resending the whole history on every call, it creates a chat once and sends the next event, fitting the context to whatever model is being called and reporting what each request cost. The customer is the application, not its user.

Bitrise Remote Dev Environments gives an agent a machine. Bitrise describes three layers: engineers attaching an editor to a cloud virtual machine; a coding agent pre-installed in that same machine on the customer's own keys; and autonomous agents that take a ticket, start a machine, open the pull request and shut down, orchestrated in parallel through an open-source MCP server. The machines are the same images and caches as its CI, which is the point — the environment the code is written in matches the one it is tested in.

MCPJam tests the thing you built for somebody else's assistant. Shipping an MCP server means shipping into a client you do not control, so MCPJam runs a server inside the major AI clients, reports whether a user's task actually succeeds, debugs the step where authentication breaks, and turns the same checks into a gate on every pull request.

NovaSynth by Noveum breaks a voice agent before its customers do. It generates synthetic callers with goals, moods, accents and behaviour — a frustrated third-time caller, someone who talks over the agent, an attempt to make it reveal its system prompt — and dials the agent over real telephony rather than replaying a mocked transcript, then scores the audio as well as the transcript. Reads against the agent's tools are replayed and writes are sandboxed, so a changed decision is exercised end to end without touching production.

Higgsfield API puts more than fifty image and video models behind one asynchronous request lifecycle, so that swapping a model is not a new integration. The customer is a developer building generative media into something, not a person making a video.

ZeroClick is the one that runs the other way. It turns an existing API or offering into an agent-purchasable service: a machine-readable storefront, support for the x402 and MPP agent-payment protocols, pricing controls, transaction analytics, an identifier for each buying agent, and settlement straight into the seller's own Stripe account. The customer is a business; the buyer it is being introduced to is a piece of software.

The two where the agent is the front door

AskDeck can be briefed by phone call, text message, email or the web — or from inside Claude, ChatGPT or Cursor through an MCP connection — and returns an editable PowerPoint and a narrated video that presents it. Project Feed includes full API, CLI and MCP access on its paid tier, alongside the human review workflow that is the actual product.

Neither is an agent product. But both treat an assistant as a place a customer already is, and build a door there. Counted with the seven, that is nine of fourteen launches touching agent plumbing in some way. The stricter seven is the number we would defend; the looser nine is the one that shows the direction.

Four of the fourteen name the Model Context Protocol on their own pages — MCPJam, Bitrise Remote Dev Environments, AskDeck and Project Feed. A year ago that would have been a specification with a following. On two days of launches it is now load-bearing plumbing for a testing company, an infrastructure company, a presentation tool and a project tracker, which is a wider spread than any of them.

The five that are ordinary business software

Jottoo turns meetings into deadline-tracked tasks without a bot joining the call. Convo puts the next move on a salesperson's screen while the prospect is still talking. Creem handles payments, tax, payouts and affiliates for people selling digital products. Figo watches competitors across search, advertising, social and the AI assistants and sends one briefing on Monday. Pitchfire for Startups reads a founder's deck, scores it against thousands of investment theses and sends the introduction itself.

All five use AI inside. None of them sells to somebody building agents. This is the group that would have looked entirely normal in 2019 with the AI removed, and it is a useful control: the interesting claim is not that AI is everywhere, which is unremarkable, but that half of a two-day sample has stopped aiming at a human operator altogether.

One of the five is worth a second look for a different reason. Figo watches what ChatGPT, Claude, Gemini and Perplexity say when a buyer asks for a recommendation, and reports where a brand is missing. That is a marketing product whose subject is an agent's opinion. It sells to a person, about a machine.

The price stopped counting seats

Look at the pricing pages rather than the taglines and the same shift shows up in the unit.

Three of the fourteen charge per person: Convo at $22.49 per user a month billed annually, Project Feed at $12 per editor a month billed annually, and Weave's platform at $50 per engineer a month. Project Feed is already hedging — viewers, clients and external reviewers are free on every plan, so only the people who change something are counted.

The rest charge for something else entirely. Creem takes 3.9% plus $0.40 per successful transaction, with no monthly fee. Bitrise RDE sells machine minutes, up to 1,440 a month per operating system, with a spend cap. Higgsfield API bills per second of generated video — $0.2057 a second for one of its models at list price, discounted at launch — and per image. AskDeck charges $29 for a finished deck, once, with no subscription. Jottoo has one flat plan at $13.99 a month. Figo prices by how many competitors are watched and how often the assistants are asked, from $49 to $249. Noveum states the principle outright: "one price per organization, not per seat", from a free tier to $599 a month. Twigg publishes no subscription price at all, only a metered model catalogue with the platform markup stated inside it. Pitchfire for Startups is free. ZeroClick and MCPJam publish no price on the pages we could read.

That is not a rounding error in how software is sold. A per-seat price is a bet that the work is done by people and that counting them measures the value. When the thing doing the work is a process that can be run a thousand times in parallel, the seat measures nothing, and the vendors have noticed before most buyers have.

The two models pulled in the same direction

Two models were added to our board in the same window, and both are odd in the same way.

Koa, announced by Salesforce with NVIDIA on 15 September, is the first model here built by an application vendor rather than a lab. Salesforce post-trained NVIDIA's Nemotron 3 Super — a 120-billion-parameter open model whose training datasets NVIDIA publishes — on synthetic enterprise workflows across more than fourteen industries, then kept the weights and runs inference inside its own infrastructure so that customer traces never become training data. Salesforce's own account says the model was deliberately trained for the case where no available tool can do what is being asked, to say so and hand off rather than call something similar-sounding. Its only performance claim is a ratio — three times fewer errors on Salesforce's own CRM benchmark, against unnamed models — with no absolute figure published, which is why our board stores no score for it.

Pareto 26.9, released by Unbiased AI on 17 September, is not a set of weights at all. It runs a mix of frontier and open models against each other on every request and keeps the best answer, returning one response for one call, and the company says the composition of the blend can change between releases. It publishes five benchmark figures and prints its losses: on four of the five, at least one of the three models it names itself scores higher, and it ties for best on the fifth. It says on the same page that no measured task cost or composite score exists for the release, and no independent evaluator had published a score for it when this was written.

Between them: a model you cannot download from a company that does not sell models, and a model that is a moving average of other people's. Both are sold on what they do inside somebody else's system rather than on what they are.

What a buyer should take from this

Three practical things.

First, if you are buying anything in this category, ask what the unit of value is and whether your bill tracks it. A per-seat contract for work an agent does in parallel will either become very expensive or very cheap, and neither outcome is one you planned.

Second, ask whether the vendor publishes what it loses on. Unbiased printing four tests where a named competitor beats it is rarer than it should be, and it is the cheapest available signal about everything else a vendor tells you. A ratio with no absolute number, like the CRM claim above, is the other kind.

Third, for anything that runs agents against your systems, ask what it does with writes. NovaSynth sandboxing writes while replaying reads is the shape of the answer you want, and it is a question worth asking of every product in the seven.

What this does not show

Two days. One platform, whose audience skews heavily towards developers and whose ranking rewards a launch-day campaign. A sample selected by our own bar for business software, with the exclusions listed above so you can disagree with them. Nothing here measures revenue, adoption, retention or whether any of these products works — only what fourteen companies chose to build and how they chose to charge for it.

One disclosure. TrustList owns listings on this launch board: our own Rutba products appear on it, most recently Rutba Office on 10 September. None of them falls in the 16-17 September window, so every count in this article is unchanged if ours are removed.

And one honest caveat about the framing. "Sold to the agent" is our categorisation, not the vendors'. Every one of the seven would say it sells to a team of humans who happen to run agents, and they would be right. The reason the distinction still matters is that the product is shaped by whatever it has to satisfy — and for seven of these fourteen, that is a process, not a person.