Make products AI agents want Make products AI agents want<br>May 18, 2026<br>“If you’re running a productivity or an infra company in 2026, humans are no longer your users. Agents are.”

The DAU/MAU ratio told you how habitual the product was for humans, where 50% and up meant you had a killer product. But that worked when your product was used by humans.

Now their agents are doing the work.

Here’re some of my observations on how to build the product agents reach for, and how to measure when they do:

Before AI agents

Measuring DAU made sense in a world where you expected users to live inside your product.

If people spent more time in it, that usually meant the product was delivering value, or it had strong enough network effects to keep pulling them back in.

In fact, many productivity tools (Notion, Linear, Figma, Slack) or developer tools (GitHub, Vercel, Datadog) adopted seat-based pricing models designed around human access:

how many people can enter the product,

what permissions they have, and

what features they can touch.

But, humans are slow to pick up new skills and learn new interfaces. To help them out, PMs and marketers obsessed over making the best onboarding experience and hyper-personalized drip campaigns.

They also made it really hard for people to churn from their products, because they already spent a significant time onboarding and learning their UI. The sunk cost fallacy is hard to fight once you’ve spent 100+ hours in a product… anyone who runs their CRM on HubSpot knows this well.

Now, luckily for us, AI is changing this narrative and many productivity tools are adapting to this new reality.

The agent-native internet

AI assistants like OpenClaw, Hermes and Vellum (disclaimer: I work here) can use any productivity/dev-tool software on your behalf. And, it looks like every company is taking notes.

The last 3 months every launch has been around agent-facing APIs, CLIs, MCPs, and connectors:

Stripe built their payments CLI for agents via Link

Notion just launched their External agents API

Cloudflare built their own CLI and basically shifted their business model

Linear has impressive Slack connectors and your AI assistant can write PRDs, close and open issues

Google published their own official CLI where agents can use their products

Figma opened the canvas to agents with their use_figma MCP.

Vercel has a really good agent-native CLI that let’s you host websites almost instantly

AI agents can file receipts, organize your inbox, create pixel perfect code from designs, monitor your finances, pull analytics from Stripe and align your team in Slack. Yes, the tech is that good and reliable .. and this is the worst that it’ll ever be.

As agents continue to get better, humans will be required for less and less work.

So, your product will need to become agent-native too.

How to build an agent-native product

To build an MVP with these agentic preferences you need to do few things well.

i. Programmatic parity: Every workflow a human can do in the UI has to be doable via API. Stripe for example has the high bar here, with near-complete coverage of dashboard actions in the API.

ii. Support for multiple agent surfaces

REST API. The minimum for an agent to work well; any agent that can make an HTTP request can use your product. Great for coding agents: SDKs in Node, Python, and Go. Language-specific wrappers around the API that save the agent from writing boilerplate for auth, retries, and types. Great for AI assistants: CLI & MCP support. Lets Claude, Cursor, OpenClaw, Vellum and other LLM assistants pick up your product. The benefit of this is that it allows the LLM to pick up your product directly as a tool call instead of it needing to write code first

iii. Docs an agent can actually read : This is the new onboarding you should care about. Agents will read your docs in 50ms and decide if they can do the task. This is where your AIO/GEO strategy becomes a mid-of-funnel technique. That means:

a working example next to every endpoint,

real request/response payloads instead of placeholders,

a single endpoint reference instead of a marketing-doc maze,

a stated versioning policy, and

an /llms.txt (or equivalent) that gives an agent the map of your docs in one file.

iv. Agent identity + safety primitives: Agents need their own token, their own rate limits, their own audit trail. Plus the must-have safety pieces:

idempotency keys (retries don’t double-charge),

dry-run or preview environments and endpoints (agents can test),

webhook signing (agents can trust events)

v. Distribution into agent ecosystems: Be where the agents already are: MCP marketplace listing, native integration into Claude Desktop, ChatGPT, Cursor and visible in agent registries (e.g. skills.sh by Vercel).

All of the above makes it very easy for agents to come back to your product and get the work done, reliably. Sadly, all of this enablement is only a temporary moat.

When the user is an agent that reads...