An administrator could act without being logged
Broadcom scored a VMware ESX logging flaw at 2.7 and CISA now lists it as exploited. GitHub's agents began writing the guidance later agents run on. The obstacle to agentic engineering turned out to be what developers think the job is.
How we Organise
The obstacle to agentic engineering was identity, not technique
Martin Fowler spent last weekend at Foocamp. Several people there described running immersion courses in agentic engineering. Each takes a week or two, away from normal work, with experienced coaches, and those coaches told Fowler the hardest part was emotional rather than technical. One said his main job was walking people to a coffee shop. They needed the time to take in what the tools do to the idea of being a developer.
We tell clients that outside help is the measure of a capability they do not yet hold.
A team that still needs help from outside has not yet acquired the capability.
Firms paying for two weeks of coaching have not acquired the capability yet. Fowler attended fewer than a tenth of the sessions. He wrote under the Chatham House rule, so the coaches are unnamed. He reports no figures.
How we Build
GitHub’s agents now write the guidance the next agent runs on
GitHub made Copilot Memory generally available. It also opened a preview of dreaming, which reads finished sessions and revises what later ones are told. The sessions generate and curate those memories themselves. Dreaming can add a memory, change one, or remove it. Where it would change a skill or instruction file that is checked in, it opens a pull request instead of editing the file.
We ask that a machine’s output be marked as the machine’s wherever it is stored.
Model-generated data is identified as such, wherever it is written.
For anything in the repository the pull request does that, and it is the half a reviewer can act on. GitHub does not say whether a memory records that a model wrote it. Both kinds sit in one store. The preferences a developer dictated sit beside the ones a session inferred. Its 93 per cent adherence figure comes from sessions GitHub selected to review, with no denominator published.
How we Assure
Broadcom scored the missing log at 2.7 and CISA now lists it as exploited
Broadcom’s advisory says a malicious administrator could use a flaw in VMware ESX to perform certain operations without them being logged. It places the flaw in the low range at 2.7. The score follows from the vector. That vector records no disclosure and almost no damage to data, because what the flaw takes away is the account of what happened. CISA added the flaw to its exploited-vulnerabilities catalogue on 28 October, three months after the advisory. The remediation date is 11 November.
We hold that oversight needs the power to halt the thing being overseen.
A reviewer who cannot stop the thing reviewed is not overseeing it.
A reviewer of that administrator could hold the power to stop them in full and still oversee nothing. The only account of what was done is the one the administrator can suppress. This is the second source this autumn to bear on that position. In September METR said its evaluation of Claude Opus 5.5 was never meant to verify compliance with any threshold, so it had no refusal to make. Broadcom does not say which operations go unlogged. It does not say how the flaw is reached, and CISA names no actor. A reader could fairly take the 2.7 as right about the damage to the data and wrong about the exposure it leaves. If you read the score, or our statement, the other way round, write to transform@dromologue.ai.