This is our public writing - what goes wrong with AI in business, how we build, and how we work with clients. Every claim has a source you can check.
Start with the one that matches what you want to know. If a section is empty, we have not written it yet.
When a public AI failure happens, we write up what went wrong and what would have stopped it. We name the pattern, not the vendor.
SectionWhy we built it this way: one install per client, logs that cannot be erased, people and agents using the same gate.
SectionA dated record of our work towards ISO 27001, SOC 2 and ISO 42001. Until we hold a certificate, we say we are working towards it.
SectionWe take someone else's article or research and test it against how we think. If they are right and we are wrong, we say so.
SectionWhy pay us when a free agent framework exists? Including when the free option is the right choice.
SectionHow a business actually changes: a named goal, the work towards it, and the person who keeps it honest.
SectionThe newest writing, at the top.
AI stalls inside companies for two human reasons: people cannot reimagine their workflows, and leaders will not manage the change. The evidence supports Claire Vo's diagnosis, published 12 August 2026, and the industry's chosen fix - sending engineers to work inside the customer's business - only reaches firms that can pay for a pod of them. Below that line the same two problems remain, and the answer has to arrive already built, with a person whose job is the change.
Buying AI seats is not the same as moving the business. What moves it is tracked, executed work toward a named goal for the company, and then for each part of it - a North Star - with AI used only where it advances that goal. Unused logins and renewed invoices are what it looks like when the goal was adoption.
Red First published nine planes for every agent - a member of staff made of software - on 11 August 2026: Soul, Knowledge, Skills, Capabilities, Listeners, Connections, Surfaces, Heartbeats, Guardrails. That list only holds if taking any one of them away, from an agent specified to use it, breaks the job or makes the output worse. This piece runs that test, and records why the nearest neighbouring plane is not a replacement.
Industry cannot agree on what an AI agent is, and Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027. This piece argues those failures trace to missing anatomy, and sets out the nine planes Red First declares for every agent: Soul, Knowledge, Skills, Capabilities, Listeners, Connections, Surfaces, Heartbeats, Guardrails. Whatever a build is called, it answers all nine questions somewhere, the only choice is whether they are answered on purpose.
Single-player AI is built for one person: one login, one memory, one set of rules belonging to whoever signed up. A business is a multiplayer game with shared information, different permissions, one goal - and single-player tools become liabilities when a company quietly starts running on them, not because they are bad but because they were not designed for a multi-player environment. The fix is not better tools or stricter bans: it is architecture, built for the business as the unit with shared memory, one gate, one goal.
Articles in a section follow the same shape, so you always know what you are getting.
When a public AI failure happens, we write up what went wrong and what would have stopped it. We name the pattern, not the vendor.
Nothing published yet. Open this section
Why we built it this way: one install per client, logs that cannot be erased, people and agents using the same gate.
1 published. Open this section
A dated record of our work towards ISO 27001, SOC 2 and ISO 42001. Until we hold a certificate, we say we are working towards it.
Nothing published yet. Open this section
We take someone else's article or research and test it against how we think. If they are right and we are wrong, we say so.
1 published. Open this section
Why pay us when a free agent framework exists? Including when the free option is the right choice.
Nothing published yet. Open this section
How a business actually changes: a named goal, the work towards it, and the person who keeps it honest.
3 published. Open this section
Before anything ships: would this survive a diligence read and a hostile journalist on the same day? If either answer is no, it does not ship. Every claim is dated, given a source, and archived.
Three tiers and hard rules. Incident records use primary sources. Named products need evidence archived on the day we cite them.
Open standardsThe words we use in public, defined once. The first time a term appears, it links here.
Open glossaryA public log with a 48-hour clock. Material errors are corrected and recorded here. Nothing to correct on what is live.
Open correctionsThe sourcing rules are public. Start there if you want the standard behind the writing.
Our standards