What we tested, and what happened.
Tools tested on our own projects: adopted, still in testing or dropped.
Nothing reaches a client until it works in our own business.
Where each tool stands8 entries
Trying1
AI-run media and marketingPromising0
Nothing here yet.
Works for us3
A legal assistant that cites Spanish lawAutomating our financial systemsA faster way for agents to test our appsReady for clients3
Jev, from TypeSafeSite down. Fixed before we knew.All business knowledge in one place, ready for AIDropped1
Adversarial multi-agent debateReal entries from our own work. Dates are being confirmed.
TryingStarted recently. Most tools stop here and get dropped.
PromisingUseful, with clear limits. Still testing.
Works for usReliable on our own business. Not yet trusted for yours.
Ready for clientsAdopted. Used in client work where it fits.
Sept 2026AdoptedClient workJev, from TypeSafe Each AI request now carries only the tools it needs: 8 instead of 19 for a banking question, decided in about 249 ms. In a client’s system 2 days after we measured it, under a week from its release.Ready for clientsAug 2026AdoptedClient workSite down. Fixed before we knew. Monitoring caught a client site going down and handed it to an AI agent. It was back up within 9 minutes, before anyone on the team had looked at it.Ready for clientsMay 2026AdoptedClient workAll business knowledge in one place, ready for AI A company’s documentation, data and processes, structured in one place so AI can use them for personalised outputs: answers, quotes, follow-ups. Running on our platform; live for a client in November.Ready for clientsAug 2026AdoptedClient workA legal assistant that cites Spanish law Answers Spanish legal and tax questions from official sources only: BOE law, all 17 autonomous communities and 41,237 binding tax rulings. Every claim carries its citation. Client access is next.Works for usDec 2025AdoptedClient workAutomating our financial systems Agents read receipts in any format and from any country, match them to bank movements and book them only when every check passes. What needs a person goes to one review queue.Works for usSept 2026AdoptedClient workA faster way for agents to test our apps Our internal testing tool: it reads each screen as text instead of sending screenshots to a large model, so an agent walks a whole app in minutes. 56 screens in 25 minutes on a real app.Works for usSept 2026TestingMedia & marketing testAI-run media and marketing Running on SEIRO first, in parts. Next: a measured test on our own ad campaigns. Ready for clients when it brings leads that turn into clients and pays for itself.TryingJun 2026DroppedClient workAdversarial multi-agent debate We set agents to critique each other’s plans and code in loops. They converge on agreement too fast, and a strong model debating a weaker one ends worse than the strong model’s first answer. Now each decision goes to the most capable model.Dropped
+Tried so far
Models and tools we’ve tried.
Each one ran on our own work before we formed a view. The entries above say what we kept and what we dropped.
Models
- Anthropic (Claude)
- OpenAI
- Google Gemini
- Mistral
- xAI Grok
- DeepSeek
- Meta Llama
- Qwen
- Kimi
- GLM
- TypeSafe Jev
- Cursor Composer
Routing and search
- OpenRouter
- Voyage AI
Voice, image and video
- Higgsfield
- ElevenLabs
- Fal
- FLUX
- Runway
- Midjourney
Coding and building
- Cursor
- Claude Code
- Lovable
- v0
- Bolt