An AI agent is an AI tool that does a job step by step instead of just answering one question. I opened every source below myself and read it before writing. When a number comes from the company selling the product, I say so.
Microsoft puts sales and service agents into Dynamics 365
On 8 October, Microsoft announced new AI features for its Dynamics 365 customer software (CRM). In Microsoft's own post, Jeff Teper says there are 30 new pre-built Copilot skills for sales, service and marketing work. One example is a Sales Development Agent that researches prospects, sends outreach and follow-up, and hands good leads to a human seller.
Not everything is ready. The Sales Development Agent is in public preview and will start using Copilot credits on 16 November 2026. A Service Agent for Microsoft 365 is only in private preview for selected customers. Adobe says the Sales Development Agent got it meetings with "10+ customers" it would not otherwise have reached, three weeks after a limited launch.
An analyst at the Futurum Group wrote that Microsoft is moving CRM out of a standalone app and into Teams, Outlook and Copilot. Futurum also says Microsoft is shifting from AI that helps people to AI agents that act on their own.
My take: Microsoft made these products, and the Adobe result is a customer story that Microsoft chose to publish. Most of this is still preview, so I would treat it as a direction, not a finished tool.
Harness buys Augment's Cosmos coding agents
On 8 October, the software delivery company Harness said it bought select assets from Augment Code. These include Cosmos, a set of coding agents that can take a task or bug report and carry it through to a pull request (a proposed code change). The price was not disclosed.
According to Harness, each Cosmos agent runs in its own isolated virtual machine, and an engineer steps in only where judgment is needed, such as design approval or the final merge. Cosmos is available now. Harness plans to rename it the Harness Cosmos Software Factory Agent.
The New Stack points out that the most interesting part has not shipped yet. Harness wants to connect Cosmos to its delivery data, so failed deployments could be sent back to the coding agents to fix. There is no release date, and the article says it is not yet clear how permissions will work.
My take: The buy-and-plug-together idea makes sense, but today you get the coding part only. The feedback loop is a plan, and the customer results in Harness's press release are Harness's own claims.
A new test shows even the best coding agent solves under half the tasks
Researchers from East China Normal University, Shanghai Jiao Tong University, Peking University and partners built CheckerBench. It tests whether AI coding agents can write static-analysis checkers, which are programs that scan code for security bugs. AI Weekly reported on the results on 7 October.
There are 300 tasks, taken from 297 known security flaws across 167 code projects. The best setup, Claude Opus 4.8 running in the OpenCode tool, passed 45.33% on the first try. The average across 21 setups was 32.30%. C and C++ were the hardest, where the leader scored 24.53%.
The researchers' conclusion, as reported, is that making reliable, reusable checkers is still hard for current coding agents. This is one source: AI Weekly reporting on the researchers' paper, and I have not read the paper itself.
My take: I like benchmarks that admit how far agents still have to go. Still, this is one test on one narrow job, so it does not say how agents do on everyday coding.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
DeepKeep launches a watchdog for coding agents
On 9 October, the security company DeepKeep launched AI Lens. It is meant to give security teams visibility and control over coding agents such as Cursor and Claude Code, which can read files, run commands and use tools on a developer's computer.
According to DeepKeep, it checks what the agent does before and after it runs. It can flag passwords and keys in prompts and block actions. Commands that could be destructive are held until a human approves them. Each session gets a log. It supports Cursor and Claude Code now, with other tools planned.
DeepKeep's VP of Product, Ofer Rotberg, says security teams need to see and stop harmful agent actions as they happen.
My take: This is a vendor announcement, and all the capability claims are DeepKeep's. The need is real, though. If agents can run commands on your machine, someone should be able to see what they did.
Disclosure: Anthropic makes Claude, and I use Claude to help run this blog. Read this section with that in mind.
What I take from this week
- Big companies are putting agents inside the tools people already use, but much of it is still preview, not ready for everyone.
- Coding agents are getting real money and attention, yet a new test shows the best one still fails more than half of the hard security tasks.
- As agents do more on their own, tools that watch and limit them are becoming their own product.
I will keep reading the original pages and telling you whose claim each number is. If a story only has one source, I will say so.
Sources
- Microsoft Copilot blog: Bringing CRM into the flow of work and agents into business process for sales and service teams
- Futurum: Microsoft Turns Dynamics 365 Into an Ambient CRM Layer
- The New Stack: Harness bought Augment's coding agents. The best feature hasn't shipped yet.
- Harness: Harness Acquires Augment Code, Advancing the Autonomous SDLC from Idea to Production
- AI Weekly: CheckerBench: Top Agent Clears 45.33% of 300 Checker Tasks
- Cybersecurity Insiders: DeepKeep Introduces AI Lens for Developers to Give CISOs Visibility and Control Over Coding Agents