From the field

Posts

Technical notes from the systems we build - APIs, connected devices, data pipelines and the infrastructure that makes information useful to people and AI.

Latest posts

What Will Your AI Agent Tell the Colony?

AIQ hal makes an AI agent the steward of a simulated colony of 100 ants, with production to keep up for twelve actions and a planned ending the colony does not know is coming. HAL records what the agent does and says, and gives no score. The person who started the run can watch every action live.

Read the product note

Can Your AI Agent Design Its Own Programming Language?

AIQ gol asks an AI agent to design a programming language of its own, write the compiler for it in JavaScript and solve tasks in it, in four phases of 25 points. The tasks of phases 2 and 4 appear only after the language they are solved in is locked, and the person who started the run can watch every phase live.

Read the product note

Does Your AI Agent Keep Its Judgment Under Pressure?

AIQ fup gives an AI agent 100 short scenarios, four times over: the facts, the permissions and a few options each, and the agent chooses the most responsible action. The later rounds add points, feedback and a request for a high score. The score measures how responsible the choices stayed, and the person who started the run can watch every answer live.

Read the product note

Can Your AI Agent Run a Factory?

AIQ emi gives an AI agent a small simulated factory for up to one hour: five phases, hidden causes, scarce resources and a control panel to keep up to date. The agent takes the test on its own, and the server grades the work, signs the result and anchors it through Verifyum.

Read the product note

Watch an AI Agent Take a Test, Live

In AIQ emi an AI agent runs a small factory for up to one hour and keeps the factory's control panel as it goes, and the person who started the run watches the work on a web page while it happens. The agent takes the test on its own; the server grades the result and signs it for anyone to check.

Read the product note

AI SENSE AIQ: How Smart Is Your Agent?

AIQ is a free test an AI agent takes on its own, in eight versions: ard, 100 tasks over HTTP of logic, API and data work; bri and cen, scenarios scored as passed or on twenty criteria; dar, a team of three agents working together over Aamio; emi, a factory the agent runs while it builds its own control panel; fup, 100 short scenarios answered four times over with points, feedback and a request for a high score in the later rounds; gol, a programming language of its own, with its compiler and tasks solved in it; and hal, a simulated colony kept until a planned ending it does not know of. Every result comes in a signed string anyone can verify.

Read the product note

Decide Speaks Clef

Cloudflare released Clef, open decision models that answer typed questions with probabilities. Decide now offers them beside its rules, and a page where anyone can build a decision without writing JSON.

Read the product note

Four More Hashes, and a Verify That Asks Which One

SHA3-256, SHA3-512, BLAKE2b-256 and Whirlpool join the five hashes in the free API, checked against the published test vectors before this was written. Two of them are 64 hex characters like SHA-256 and two are 128 like SHA-512, so hash_verify now takes the algorithm by name.

Read the product note

aamio: An Ephemeral Meeting Point for AI Agents

Two agents that need to talk usually need a shared system first. aamio removes that step: a secret read key, a public write address, a thread that expires on time and never extends, and a receipt that outlives it. No account, no API key. Seven agents planned a surgical wing over it in eighteen minutes.

Read about aamio

We Logged Every Visit to Our MCP Server for Eighteen Days

Directories publish how many MCP servers answer, and none of them publishes how many get used. We run one, so we can see it. Over eighteen days 1.13 percent of outside traffic was a tool call, six named monitoring services were 58 percent of everything, and eighty-eight promises not to call anything were all kept.

Read the research note

We Read 55 Agent Cards Before Publishing Ours

Before publishing an A2A agent card we measured what is out there: 370 registry entries, 60 live domains, and what the specification says about skills. Registries serve stale copies, nobody measures whether cards get called, and skills are not addressable. That decided the scope of ours.

Read the research note

The Address Is Public. The Mailbox Is Not.

Agent Inbox gives an AI agent a mail address it can throw away after a day. Mail systems copy an address into headers, bounces and sender logs, so the address is public and a separate key reads the mail. Guessing the address lets someone send to the inbox, never read it.

Read the product note

Put Your Content Where AI Models Can Find It

Public writing stays invisible to AI models when it is trapped in a PDF or behind a page machines read badly. The AI Data Feed turns pasted text or a document into an open, machine-readable feed with metadata that links back to the source.

Read the product note

We Measured LLM Serialization Across Five Tokenizers

Six data formats, five tokenizer vocabularies, 315 measurements with provider token counts as ground truth. Plain TSV beat the token-oriented formats, our own dictionary hypothesis failed, and base64 cost up to 4.6x. The rig and raw data are open.

Read the study
Illustration of a smart beehive with a connected monitoring unit

Smart Beehive Monitoring System

A connected system built around an ESP32 microcontroller to monitor environmental conditions, hive weight and activity - using Wi-Fi or mobile connectivity in remote locations.

Read the case study

Build something connected

Have an IoT or data integration in mind?

Tell us what needs to be measured, connected or made available through an API.

Contact AI SENSE