The agent landscape, read backwards from the work.
Every map of this market starts with vendors. Three hundred-plus companies across eighteen categories, each homepage promising to run your business. This one starts from the other end, with the task a person actually has, and asks which layer of the stack is even involved. Below: 140 tools, twelve layers, and an honest account of where our own product sits.
A supply problem the market keeps treating as a demand problem.
Capability is not the bottleneck any more. Naming the work is.
Everything, for everyone
Read ten homepages in this directory and you will find the same sentence. Agents that run your business. The category is genuinely differentiated on the inside and almost identical on the outside.
The buyers are not founders
A founder reads a pricing page and infers a workflow. A bookkeeper, a paralegal or a store manager cannot. Most people have never seen their own job written down as a list of tasks, let alone scored.
The failures are scoping failures
Gartner expects four in ten agentic projects to be cancelled by 2027. MIT found almost no measurable return across generative pilots. Very little of that is the model being incapable. It is nobody naming the job first.
The stack, ordered by distance from the human.
Most market maps group by company type. This one groups by how far a layer sits from the person with the task, because that is what decides who has to understand it.
The directory.
Entry price as published by the vendor, checked July 2026. Blank means the vendor does not publish one. This market reprices constantly, so treat every number as a starting point rather than a quote.
Generalist agents
Vertical agents
Enterprise platforms
Builders
Frameworks
Tools and auth
Web, data and docs
Memory and knowledge
Runtime and sandboxes
Eval and observability
Models and routing
Discovery layer
Method. Figures are vendor-published or press-reported and have not been independently verified. Funding, revenue and star counts move weekly. Nothing here is a paid placement. Where a tool spans layers it is filed under the one its own homepage leads with.
Where our own product stands, stated plainly.
A directory that ranks tools and quietly owns one is worth nothing. So here is the position, including the part that does not flatter us.
Where it cannot win
- Composio lists over 1,000 toolkits. Nango is around 900. Pipedream connects roughly 2,700 APIs. Make lists 3,000-plus apps. Lindy claims 5,000-plus integrations.
- Keimodel ships one read connector, for Shopify, plus three delivery channels.
- That gap is three orders of magnitude and it is bought with capital and years, not cleverness.
- Treating it as a roadmap item is how a small team spends two years losing a race it already lost.
Where it can win
- Routing across 200-plus models with a visible receipt for why each was picked. Almost every builder in this directory hides the model choice or hard-codes it.
- Teams where each member runs on a different model and hands off. Sold elsewhere as an enterprise feature.
- Most useful work needs no integration at all. Research, drafting, analysis, planning and reporting all produce a deliverable from a prompt and a web search.
- It is the only tool on this page attached to a demand engine that knows what the visitor’s job is before they arrive.
What it therefore is
- Keimodel runs a real sample of your task in the page, free, before you have signed up for anything.
- Anything needing a live system of record gets routed out to the vendors above, and that routing is the business.
- Keimodel is never ranked against them here. It is the try-it surface, labelled as ours.
- Which tasks people try, and which routed clicks convert, is the signal for the only connectors worth ever building.
What we are building toward.
The short version: everyone here sells a hammer to people who cannot name their nail. The index of nails is the asset.
The unit is the task, not the tool.
Every directory in this market has the vendor as its atom, which is why they all converge on the same seventy alphabetised categories and the same six advertising slots. They start from supply because supply is what emails them.
Our atom is the task. 1,016 occupations and roughly 19,000 task statements from the public-domain O*NET database, each scored for how much of its output an agent can actually produce, each mapped to the tools that could produce it. A vendor list is copyable in a weekend. A calibrated map of work to capability is not.
Directories index who is selling. We index what needs doing.
We own the entry point nobody wants.
Risk quizzes tell you your job is 43% exposed and stop there. Agent builders open a blank canvas and assume you already know what to build. The distance between those two screens is where every abandoned pilot lives, and it is unclaimed because it is unglamorous and needs a labour dataset rather than a model.
Three page types scale programmatically off data we already hold: the occupation, the individual task, and the tool. The fourth type, task crossed with tool, is where the commercial intent is.
The demo is the moat.
No comparison site can execute. We can. A visitor sees their job broken into tasks, then watches one of those tasks actually run, on a real model, in the page, before signing up for anything. That is not a feature a directory can add later. It requires owning a runtime, which is what Keimodel is for.
This also resolves the conflict of interest cleanly. Keimodel does not compete in the rankings. It is the bench where you try the work, and the ranked vendors are where you go to do it at scale.
Three revenue layers, in the order they become possible.
First, referral on routed clicks. It needs no business development and costs nothing to turn on. Second, placement against specific task clusters, which is worth far more than a banner because the intent is legible: this visitor is a bookkeeper who just learned that invoice reconciliation is 88% automatable. Third, and the real one, the demand data.
Which tasks, in which occupations, people are trying to automate this month is a dataset nobody else can assemble. Every vendor sees only its own funnel. We see the question before the tool is chosen. That is worth more to Sierra or Lindy than a logo slot, and it compounds.
The risk, named before someone else names it.
This is a media and marketplace hybrid, and both halves are exposed. The tool layer will consolidate, and AI answer engines are already absorbing the traffic that classic directories lived on.
The hedge is structural rather than hopeful. Task-level structured data is the format answer engines cite rather than replace, and the task graph survives consolidation intact. If the vendor list collapses from 300 companies to 12, the question of what a paralegal’s day is made of does not change at all. Only the right-hand column does.