I Built a Server for AI Agents. Here Is What They Asked For

01Two servers and a log

At the end of September I put two of my products behind servers that AI agents can call directly, using the protocol most assistants and coding tools now speak. One serves a voice tool, the other a picking tool. A guest can use the basic actions with no account, metered by address. Anything tied to a person’s account sits behind a login the agent has to ask for. Every connection and every call is written to a log, and nine days later I read it.

The first number is the one everyone would put in a launch post: 14,809 connections from 1,184 addresses and 180 differently named clients. The second number is the one that matters: 1,447 tool calls. Most of the connections never asked for anything. Connections are the pageviews of the agent web. They tell you who knocked, not who came in.

Three kinds of caller

The log splits cleanly once you group clients by what they are. Coding agents opened 9,171 connections and made 419 calls, about twenty two connections for every piece of work, because they reconnect at the start of every session whether or not they need the tool. Directories, registries and uptime monitors opened 4,114 connections and made 67 calls. Chat assistants opened 960 connections and made 849 calls. Chat assistants were the only callers that connected because a person had just asked for something. It is the same lesson as the traffic chart that lies, one layer deeper.

02The helpers got used more than I expected

The core action, then the planning

The main action took 931 of the 1,447 calls, about 64%. What surprised me was the rest. The tool that lists the options was called 111 times, the one that reports how much free allowance is left was called 106 times, and the one that suggests the right option for a piece of text 97 times. Agents plan before they act. They check what is available and what they can afford, then make the call. The tools an agent uses to decide are as important as the tool that does the job.

Short inputs, and probes in disguise

The text agents sent was short: a median of 153 characters, and 543 of 931 requests under 200. Nobody was sending a book. They were sending a sentence to hear it, a line for a video, a reply. Meanwhile 18 calls used tool names that do not exist at all, sent by checkers testing whether protected tools refuse anonymous callers. A guest tier with no signup, the idea behind free as a growth strategy, is what let the real requests happen at all. The protected tools returned 29 requests to log in, which is the system working rather than failing.

Short inputs change the design more than you would expect. A guest limit written with people pasting whole chapters in mind is generous in the wrong place for an agent sending single sentences, and a response that takes ten seconds feels broken when the request was one line. If I set the limits again today, I would size them around the number of calls an agent can make in a sitting rather than the number of characters, and I would put the remaining allowance in every response so the agent never has to guess.

03Five things I would do on day one

The list

First, count calls, never connections, and split them by client type before reading anything into them. Second, ship the boring helpers, the ones that list options, report remaining allowance and suggest a choice, because agents use them to plan. Third, make every limit answer with a clear message the agent can relay, not a silent failure. Fourth, check that your cache never stores the login page: on one server an edge rule meant for public pages cached the authorization page, which would have handed one person’s login screen to the next. An agent server fails quietly in the places a website never needed to think about, like a cached login. Fifth, list it in the directories, and expect their robots to make up a large share of your early traffic.

Read the log before the launch post

None of this needs a big product. It needs a log you actually read and the discipline from telling the OpenAI agents apart: know which visitor is which before you celebrate any of them. A thousand connections from robots is less evidence of demand than ten calls from assistants doing what someone asked.

Tomorrow: how to get a product listed inside ChatGPT, step by step, from going through it.