Not subscribed? Sign up to get it in your inbox every week.

Somewhere mid-task, every coding agent used to hit the same wall.

It could write the whole app: the frontend, the logic, the tests. Then it needed a database, auth, and somewhere to deploy, and the work stopped dead while a human went off to create a cloud account, click through a dashboard, and paste keys back into the session. The most capable worker in the room, waiting on its owner to run an errand at the company store.

A YC company called InsForge exists to delete that wall, and its tagline says the quiet part in four words: the agent-native AWS. The whole backend (Postgres, auth, file storage, edge functions, hosting, a model gateway) is exposed through MCP, so the agent provisions what it needs and keeps building. Dashboards exist for humans who want to watch; the customer doing the work is the agent. In June I wrote that the buyer of your product is going to be an agent. I undersold it. Agents are starting to acquire their own tools, mid-task, the way you'd grab a wrench.

Watch what actually changed. Getting a new tool used to be a human ceremony: evaluate options, sit through a demo, get budget approval, create the account, configure the thing. For an agent, tool acquisition collapses into the work itself. Need a database? Provision one. Need to ship? Deploy. The evaluation happens by reading docs, the configuration happens through the API, and the whole loop runs at task speed instead of meeting speed.

Today one human click survives in that loop: the purchase. Somebody's card is on file, and somebody said yes once, up front. But that click is already dissolving at both ends. Payment rails built for agent transactions are live (the Coinbase-led x402 protocol has reportedly processed over a hundred million agent transactions in its first months), and Anthropic's Project Deal points at full agent-led marketplace purchases. The direction is a spending budget delegated to the agent: here is your monthly allowance, here are the categories, keep receipts. The card becomes an allowance, and the purchase becomes one more step the agent takes without stopping.

InsForge is the version of this for coding agents, and coding agents got it first for a reason: infrastructure is the tool they need most often, mid-task, at 3am, with nobody awake to click. But every tool category has the same wall in it somewhere. Data enrichment, transcription, monitoring, a design asset, a phone number, compute for a one-off job: any of it can be a thing an agent needs for twenty minutes to finish an assignment. All of it becomes purchasable by the worker that needs it.

Here is why this stops being a curiosity and becomes a market. On June 3, Cloudflare's CEO posted that bots had passed human web traffic for the first time: 57.5% automated versus 42.5% human across the sites Cloudflare can see. He had predicted that flip for 2027. It arrived eighteen months early, because agentic traffic grew roughly 7,851% year over year.

A human running an errand visits five sites; an agent running the same errand visits hundreds. Agents work in parallel, around the clock, and spawn other agents when a task splits. My expectation is that this ends with orders of magnitude more agents on the internet than humans, and the crossover we just watched is the early, gentle part of that curve. Every one of those agents is a potential acquirer of tools, which means the number of buyers on the internet is about to outgrow the number of people on earth.

For an operator this lands twice, once on each side.

On the buy side: your agents are getting budgets whether you plan it or not, so plan it. Write the spending policy before the first surprise invoice instead of after. It fits in a paragraph: what an agent may buy on its own (usage-priced tools under a monthly cap), what needs a human approval (anything with a contract term), where the receipts land (a channel a human actually reads), and which card it draws on. Your expense policy has a meal per diem in it today and nothing for the compute an agent rented overnight. That gap is a week away from being your problem.

On the sell side: if you sell software, some of your next thousand customers have no eyes. An agent evaluating your product reads the docs, tries the API, and decides in minutes, which means the demo call is a wall, per-seat pricing is a category error for a buyer that exists for twenty minutes, and your beautiful landing page is invisible. The vendors that win agent customers will look like InsForge looks to a coding agent: machine-readable pricing, instant programmatic signup, usage-based billing, docs that read like an operating manual. In June this was my speculation. InsForge, x402, and a traffic chart that crossed 50% are what it looks like as it starts coming true.

Two moves, one per side of the ledger.

Buy side: write the agent spending policy this week, the one-paragraph version above, and put a cap on it. The number can be small; $200 a month buys an agent a lot of twenty-minute tools. What matters is that the allowance and the receipts channel exist before the spending does.

Sell side: run your own funnel the way an agent would. Start from your docs, and see whether a machine could get from there to a working API key and a first successful call without a human intervening or a sales rep emailing back. Time it. That number is your conversion rate for the fastest-growing buyer population on the internet.

The buyers are multiplying. Make sure both your wallet and your front door are ready for them.

Till next time,
Chris