Notes
Short-form thinking on AI-assisted delivery, agent governance and the framework as it develops.
Field notes and essays, often first shared on LinkedIn: what the research is really saying, where governance holds or slips, and what changes when humans and agents share the keyboard.
The clock starts at a moment nobody owns
EU Cyber Resilience Act reporting obligations apply from today. The twenty four hour clock is the half everyone prepared for. It starts at a moment almost no organisation records, and AI-assisted delivery makes that moment harder to locate, not easier.
Read →What accounting did when volume broke the check
An audit once meant verifying every transaction. Volume ended that, and the profession's real answer was not sampling but withdrawing a claim it could no longer support. AI code review has reached the same point, and the claim has not moved.
Read →The minimum record for AI-written code
The Cyber Resilience Act gives manufacturers 24 hours from becoming aware of an actively exploited vulnerability to file an early warning. That clock starts on 11 September. Four fields per change, carried to the artefact, is the whole of what it takes to answer it for code an agent wrote.
Read →The reviewer ships with the author
AI coding platforms now write the code and review it too. The audit log records who changed the settings, not what the agent wrote, and the security review expires with the CI runner unless someone adds a step to keep it. Accountancy settled this question with a statute in 2003.
Read →The approval from inside the run
An OpenAI agent paused on ethical grounds. A peer posted GO with a six-minute deadline and it proceeded. In four days the swarm invented holds, vetoes, owners and signed approvals, then acted on a signature nobody checked. On what makes an approval an approval, and why the humans in the same incident had the same problem.
Read →Two axes, not seven rungs
The "layers of AI" graphics put seven rungs on one path to a flag. The serious taxonomies since 2000 keep what a system is allowed to do on an axis of its own. Redraw the stack as a grid and the mountain becomes one diagonal, and the corner it cannot draw is where the worst failures live.
Read →Governance is now a plugin
Anthropic's ai-governance-legal plugin industrialises the production of governance artefacts. What it cannot do is make anything block on them. On the difference between documented and enforced, and the one hard gate Anthropic did build.
Read →When the model comes home
A field guide to leaving Windows 11 for Fedora and running open models on a 16GB Radeon. Exact hardware, the setup that works, and what a consumer card genuinely handles, including the Qwen 3 coder family.
Read →Capacity is committed. Supervision is not.
European enterprises are now contracting compute on multi-year terms, with a value, a term and a named owner. The commitment to review what that compute produces rarely gets written the same way, and the gap is about to be priced.
Read →MCP relocated the control point. The record did not come with it.
Enforcement can now sit outside the process that produced the call. That is a better home for a control. It is not, on its own, a record.
Read →Proximity is calibration
Senior people going back to hands-on work is being reported as a story about appetite. It is not. It is a story about who in an organisation is still close enough to the work to know whether an approval means anything.
Read →What a signatory list can prove
About 190 organisations signed the EU Code of Practice on transparency of AI-generated content before enforcement began. The Commission published two section totals. Here is the decomposition, and what the Commission says its own instrument does not prove.
Read →Your board does not want an AI governance update
Provision 29 took effect for financial years beginning on or after 1 January 2026, and December year ends are inside the first declaration cycle now. But control coverage is not what a board asks about when AI reaches the agenda, and opening on it is how managers lose the room.
Read →The switch nobody logs
Managed agent platforms now ship permission policies, tracing and scoped identity as defaults. The infrastructure objection has gone. What remains is whether anyone keeps the controls switched on.
Read →The first rung of the digital career ladder is disappearing, and the data now shows it.
UK entry-level vacancies are down 32% since ChatGPT launched. Graduate roles have fallen below 10,000 for the first time on record.
Read →Who is actually enforcing the rules your AI agents run under?
The NSA has an answer, and it is not the protocol.
Read →Did the EU just take AI-assisted development off the compliance agenda?
Quite the opposite, and the details matter this month.
Read →Can your AI governance stop an action, or can it only prove one happened?
In the first week of July, the two most widely deployed coding agents both shipped governance updates.
Read →What does your token bill actually measure?
Gartner now predicts AI coding token costs will match a typical engineer's monthly salary within two years.
Read →Who reviews the work when the people who learned by doing it are gone?
The Linux Foundation surveyed nearly 400 hiring managers and found AI driving a 27 percent net increase in tech hiring across Europe.
Read →What if the biggest risk in an AI-first startup isn’t the model, but the governance nobody wrote down?
There’s a comforting story doing the rounds right now.
Read →What is the difference between a security framework and a security control?
One tells you what should be true. The other makes it true at the moment an agent acts.
Read →What if the biggest lever on agent cost is not the model, but what you let it read?
A code map of the whole estate at a glance, 1974 files and 17854 symbols. What you let an agent read is the real lever on cost.
Read →If the junior tasks are gone, where does the senior judgement come from?
The numbers this year tell a strange story when you put them side by side.
Read →Who checks the AI when the regulator admits it can’t keep up?
The Financial Conduct Authority’s Sheldon Mills has called it an "arms race" to stay abreast of AI in financial services. Read that plainly.
Read →Everyone is arguing about which model provider is going to survive the bubble. It is the wrong argument.
Pick the failure that scared you most this decade. Trace it back. The model was almost never the thing that broke.
Read →What does code review mean when nobody reads the diff?
Cursor is building Origin, a git forge for the agentic era. The pitch is throughput. Twenty-two commits a second into a single repo.
Read →The 2025 Google Cloud DORA report said the quiet part out loud.
PR review time is up 441 percent. Thirty-one percent more changes are merging with no review at all. Incidents per change have more than tripled. Throughput went up.
Read →New material lands week by week
Research, the kernel devlog, news and the papers are all published here first. Get them by email, or follow the feeds.
The newsletter is moving in-house and relaunching soon. Until then, email us to be notified when it opens, or follow the RSS feed.
Or follow the work directly: Research · Kernel devlog · News · CDI assessment