Microsoft and Hugging Face built a benchmark that grades agents on the database state they leave behind. In a 12-model ablation, 67.24% of failed runs still ended with a clean tool call and no error. Here is what the executable checks measure, which numbers cover only the ablation, and what to assert before you commit agent work.
A Harvard job market paper built on Jellyfish data from 718 firms finds coding agents raise lines of code by 30%, commits by 20% and pull requests by 23%, while issues and epics resolved show no statistically significant change. Pull request review time rises by 49% and the change-request share nearly doubles.
GitHub's Octoverse 2026 reports that more than 40% of pull requests in September began with a coding agent's first draft. That counts who opened the first draft, not who wrote the final code, and the report's own methodology notes make the year-over-year comparison harder than the headline suggests.
GitHub says an attribution bug left some Copilot IDE-agent activity out of usage reports and inflated CLI metrics. Here is how to interpret the gaps and check client fixes.
PostgreSQL 19 shed several planned features during its beta cycle. Peter Eisentraut counts about 8 to 10 significant reverts, depending on how you count, and describes 'a list of like forty defect reports' where each one carries a fixed overhead until 'you run out of time'. His explanation for the pattern is his own opinion: LLM-assisted code review surfaces defects faster than volunteer reviewers can fix them, and the effective quality bar rose after the features were written. The project gives its own reasons for the removals, citing reliability, a predictable schedule and unresolved design questions, and it does not attribute them to AI review. No independent defect count exists. PostgreSQL 19 is due on 19 November 2026 after a second release candidate on 12 November, and the reverted features are expected to return in PostgreSQL 20.
What should an agent be allowed to do?
Follow the permissions from sign-in through tool use. Look for approval before sensitive actions and an audit trail that records what happened.
Google says its enterprise agent runs under its own Workspace identity, logs every action under its own name, and can be paused by a per-project spend cap. Release status, retention and pricing are unstated, and every customer figure is vendor-supplied.
1Password for Agents, launched 22 October, signs a personal AI agent into sites while your password stays in the vault and you approve the sign-in on your own device. What you approve is a sign-in, not everything the agent does once it is inside.
An independent researcher showed that instructions planted in content one AI agent reads can pass through ordinary delegation into a second agent that trusts the first, and that five organisations separately confirmed and fixed SSRF bugs in their MCP servers. There is no evidence of exploitation in the wild and the acknowledged issues are fixed, while researchers disagree about whether the pattern is a new attack class or plain indirect prompt injection. If your assistants are wired into internal tools, the trust between them is part of your attack surface.
Microsoft's security leadership published a doctrine on 19 October 2026: agents should not inherit a person's standing access, and the model cannot be its own security boundary. The principle is sound and portable. The monitoring is partly in preview, extending Entra's agent security needs a Microsoft Agent 365 licence, and the company selling it holds stakes on both sides.
Anthropic donated the Agent Skills standard to the Linux Foundation's Agentic AI Foundation on 22 October 2026, transferring the name and logo, the GitHub organisation and the agentskills.io domain. The foundation has published governance and ten projects, but the security guidance Anthropic has proposed for people who write and install skills remains unpublished, and the format specification defines no trust, permission, signing or review model for a skill an agent will run.
Apple says it will introduce additional controls so that granting macOS Full Disk Access requires 'very explicit user action.' That is stronger consent, not a new block. Here is what the permission exposes, why an AI agent holding it differs from a backup app, and what the Meta Muse episode leaves unresolved.
How do agents act on a customer’s behalf?
A shopping or banking workflow can involve separate decisions about who can sign in, what they may buy, and who approves payment.
American Express published an Intent Intelligence Framework on 13 October 2026 that asks a cardholder to approve an agent's shopping instructions before a payment credential is issued. The control sits in the network and the credential, unlike the agent-side budget cap in our earlier piece on Google's enterprise agent. No live deployment was found, and the Create Intent API's public status is unconfirmed.
Walmart published seven rules for AI shopping agents alongside per-agent profiles and a beta gateway. The text covers approval, spending limits and blocking, and leaves admission, partner access and dispute routes unexplained.
Jamie Dimon said JPMorgan will not let 'just any outside agent' in, and a spokeswoman confirmed the bank blocks third-party AI agents from acting for retail customers. What has surfaced is an on-record posture rather than a written policy, and liability, enforcement and open banking stay unresolved.
Spotify's Q3 2026 deck documents 788 million monthly active users, 306 million Premium subscribers and EUR 4,969 million in revenue, and Spotify has announced agent integrations with Meta's Muse, ChatGPT and Claude. It discloses no agent-attributable revenue, retention or listening share, and the company's published developer terms let it limit or withdraw third-party access.