Developer tools get most of the attention in the agent conversation, but the recent Show HN queue is full of products pointed somewhere else: at QA teams, at market researchers, at small contractors, at people who want a thirty-second video. What they share is that the agent is hidden. Nobody is typing prompts; the system does one job and hands back a result you can check.
Testing what an agent does to your service
If your product exposes an API, an MCP server or just a website, an assistant may soon be the thing using it. That is a compatibility surface you cannot test with unit tests.
Consumer Agent Testing Tool (CATT) is built for exactly that. You run a scripted or model-driven agent against an MCP server, an MCP app, a website or an HTTP API, then check what the service actually did on its side rather than reading the agent's transcript. Results are kept as sealed evidence so runs can be compared across builds, and Muse and Instinct are supported as surrogate emulation targets behind a shared connector testing contract. The project is careful about what a pass means: CATT is a surrogate that approximates an assistant's host, not its private implementation, so a passing test does not certify compatibility. That honesty is rare and useful.
Simulating a population instead of guessing
Surveys are slow and expensive, and asking a model "would people buy this" produces confident nonsense.
Jahan takes a more defensible route. It builds a population of LLM personas, each grounded in a real survey respondent sampled from a one-million-row dataset rather than invented by a prompt, places them on a social network, and lets them meet a product on an X-like feed, a Reddit-like forum and through word of mouth. They react, talk, change their minds and answer survey waves about whether they would buy. Two design choices make the output auditable: every turn is written to an append-only trace so any run can be replayed or resumed, and prompts are rebuilt from the record and hash-checked rather than stored. Intent is scored from what personas say in their own words, not requested as a number.
Two vertical tools that read the government
Government data is public, enormous and formatted for lawyers. Two tools this month turn it into something a person can act on.
Firmenza is an AI assistant for SAM.gov and federal contracting. It monitors your registration, finds relevant opportunities, explains what a government section asks and checks the format of what you type, emails before your registration expires and prepares renewal answers in a worksheet. A free health check shows what SAM.gov lists publicly for your UEI or legal name. It states clearly that it is software, not a lawyer and not part of the government, which is the right disclaimer for the category.
ORION Catalyst Calendar is narrower and more interesting. Biotech calendars usually give you a date. ORION gives you the sentence behind it: every readout or FDA date links to the SEC filing it came from, and when a company changes its wording both versions are kept. There are no forecasts. It is free with no signup and covers readouts, FDA dates and company-stated windows by ticker, company or program. For anyone who has watched a company quietly move a date from "before the end of the year" to a specific week, the revision history is the product.
Knowledge a team can actually keep current
Internal knowledge bases decay because nobody updates them. Proposals are a better model than edits.
this+that applies that idea to a company brain. Operational knowledge lives in one place, updates arrive as proposals a person approves rather than as silent edits, DoBox turns requests arriving in your messages into tasks with an owner, and contacts assemble themselves from your own channels. An MCP server exposes the same brain to assistants you already use, which is what stops the knowledge from being trapped in a wiki nobody opens.
Three generators that do one thing
The consumer end of the queue is all narrow templates, and that narrowness is the selling point.
A2O LITE AI turns two photos into a NEON-style dance duo video. Photo 1 maps to the foreground dancer with white sunglasses in the opening shot, photo 2 to the dancer behind at the start, and the template is fixed. A short 480p preview costs far fewer credits than the full render, and the credit cost is shown before you commit.
OraStage does the same shape for the Hotel Lobby format: two photos become a rap duo on an orange stage at 9:16 or 16:9, in 480p or 720p, at five, ten or fifteen seconds. The beat is generated rather than licensed, so there is no track to clear. The first video is free with ten credits and failed renders are refunded, which removes most of the risk of trying it.
BeatsFlow is the audio equivalent and the most useful of the three. It generates instrumental beats from a text prompt or from reference audio, isolates four stems (drums, bass, melody and FX) and clones a reference style, with output aimed at finishing in FL Studio, Ableton or Logic Pro. Fifteen free studio credits need no credit card. The stem split is what separates it from a toy: you get parts you can actually arrange.
None of these eight is trying to be a general assistant. Each picks one job, publishes what it can and cannot do, and lets you verify the result. That is a healthier direction than the last two years of everything-in-one-box.
