Two kinds of work, and the things we’ve made.

Given a topic, the stack searches the literature and the web, reads what it finds, and judges each source on its recency, its academic provenance and much else. Then it drafts a referenced review. A panel of AI reviewers pushes back: is it deep enough, and does each source say what the draft claims? A citation that points to nothing on file stops the draft.

The State of Play exists because there was no good way to keep up with AI quickly. Each night it takes the practices due for review and weighs fresh evidence by how far it can be trusted. A vendor describing its own product is context, not proof. Before promoting a practice it must make the strongest case against doing so, and demotions are always a person’s call.

The same stack feeds our other research newsletters, Gaia’s Web among them. It does the reading. We still choose the questions.

Claims used to start with interviews: engineers asked to recall, long after the event, what they had tried and what had failed. That took their time, and qualifying work slipped through because nobody remembered it.

The agents, built with Perly Consulting, read what a firm already keeps: its code history, its tickets, its design notes. They weigh each piece of work against the CRA’s three tests, and against Perly’s 25-year record of real SR&ED claims and how each one was assessed. Then they draft the claim narrative within the form’s word limits. Perly’s specialists still make the judgement calls.

The records remember more than people do, and the agents turn up work the interviews missed.

Agents did the reading first: fifteen research reports and a dossier on each beach. Three rival designs were scored, and the quickest route to something useful won. It was live three days after the first conversation.

Our model starts where the apps stop. It works from the physics of the local seabed, taken from sonar surveys of each bay, and every hour it works out the swell at each peak on four West Cornwall beaches from nine open data sources. It gives face heights, how sure it is and why. Live buoys check it every half hour, and the model learns over time from that measured data, and from the surf logs kept by the people who use the app.

Most small firms have a CRM. Few keep it current, because keeping it current means typing. Ours was no different: the real figures lived in a to-do list.

Now the partners just talk: on calls, in chat, by email or by phoning an agent. Three agents keep the records. They pick up call transcripts every half-hour and sweep the inboxes each morning. Routine updates go straight in; a new contact or a payment waits for a partner’s word.

The morning brief leads with one task. A read-only dashboard shows the pipeline, highlighting who has gone quiet, and models cashflow, highlighting unpaid invoices. Only approved fields ever leave the server.

There’s no architect: the house is entirely self-designed, and Dave will be the one building it. So the whole design lives in code. Plain text files describe every layer and member, and the structure, energy model, carbon and materials list are all worked out from them. We made the calls; the AI wrote the code and ran the numbers.

We wanted small arches, tied at floor level, and roof joists short enough to fit by hand. The model agreed, with a catch: a tie alone would let the frame sway five times the limit, so the scheme adds stiff piers and a braced floor. The south overhang keeps direct sun off the glass at midsummer and lets nearly three-quarters through in midwinter.