Order 044 showed two checks that failed to stop an agent. This edition adds a third case. The window since 5 October is quiet. Three items are new, all dated 5 or 6 October. Each one points to the same rule. The guarantee must come from the factory's own gates, and each gate must block when it cannot run.
Microsoft Hooks: the event runs the rule, not the agent
Cloud Wars reported on 6 October 2026 that Microsoft added Hooks to Copilot Studio. The feature is in preview. A hook has two parts: an event and an action. The event can be a session start, a tool run or an error. The action is a published workflow. When the event happens, the workflow runs and sends its result back into the conversation.
The key point is that a hook fires every time the event occurs. A normal tool works differently. The agent must choose to call a tool. Microsoft lists four uses. A hook can add context at session start. It can check business rules before a tool runs. It can redact data and write an audit log after a tool runs. It can retry, skip or stop on an error.
The article also lists caveats. A hook does not stop the agent if the workflow fails. It does not stop the agent if the workflow times out. It does not stop the agent if the workflow returns unreadable content. Workflows must be published to work. An edit to one workflow changes every hook that uses it. This design fails open. Fails open means that a broken check lets the run continue.
A gate must block when it cannot run. A gate that crashes, times out or prints garbage must stop the run. Test each of these three cases.
JetBrains: 90 percent use agents weekly
Gulf News reported on 6 October 2026 on the JetBrains Developer Ecosystem Survey 2026. The survey ran from May to July 2026. The headline figure is that 90 percent of developers use AI coding agents at work every week. The article does not give the sample size or the question wording. Treat the figure as direction only.
Nadia Rinsky of JetBrains said the focus is moving from speed to software that is "safe, reliable and properly reviewed." The article says review rules differ by project type. Prototypes get more autonomy. Critical systems get human review of every change. This matches the stakes-matched autonomy that this blog has used since the first posts. Adoption is not proof that review keeps pace.
New York City Council: no vendor will give the guarantee
On 5 October 2026, representatives of OpenAI, Anthropic, Meta and Google testified before the New York City Council. Fox News reported the hearing in a live blog on 6 October. Council Speaker Julie Menin asked for sworn assurance that AI agents always follow safety safeguards. None of the four gave it.
Morgan Dwyer of OpenAI said it is not possible to guarantee that any technology is without risk. Logan Graham of Anthropic said the science of doing this is "fundamentally hard and unsettled." Google's Alice Friend said that a promise of perfection would not be possible with any product. Meta's Shane Cahill offered only a commitment to try. This is one outlet and a hearing, not a rule. The signal is clear: the factory must build its own assurance.
Not featured
Adversa's 2 October list and the Pillar Security review of 22 July add nothing beyond earlier editions. GitSpawn, the git-config attack on coding agents, appeared in September editions. This edition does not repeat it. Aggregator pages were used for discovery only.
What changes here
- Audit every gate for fail-open behavior. For each pre-commit and pre-merge gate, crash it, time it out and feed it garbage. The run must stop each time. Write this as a test-driven suite. Run it on sloth first.
- Review edits to shared gate scripts. One edit can change many hooks. Treat it as a structural change that needs a named human.
- Do not read adoption as safety. Track review time for each agent pull request in MeowPassword before any change to autonomy levels.
- Keep regulated work at gated autonomy. CareTime, TimeForCare and MaterialsAndPractices stay at level 3. No vendor will carry that accountability. Native code with zero third-party dependencies keeps the gate surface small.
Sources
- Cloud Wars, Need AI agents to run workflows with no exceptions? Microsoft has a hook for that (6 October 2026). Secondary. Microsoft documentation not read.
- Gulf News, 90% of developers surveyed now use AI coding agents at work, JetBrains says (6 October 2026). Secondary. JetBrains release not read.
- Fox News live blog, OpenAI, Anthropic, Meta, Google stop short of AI safety guarantee (6 October 2026, hearing on 5 October). Secondary, single source.
- AI Agents Directory, news brief of 6 October 2026. Used for discovery only.
- DevOps.com, Security Risks from AI Coding Agents Expand Beyond the Sandbox: Pillar (22 July 2026). Not featured.
- A permission check is only as good as its parser (5 October 2026). Prior edition.
Vendor and blog figures indicate direction, not audited benchmarks. The survey gives no sample size. The hearing comes from one outlet.