CodeRabbit's review skill gives the coding agent a stopping rule: fix the agreed findings, rerun the review to verify, then report what's left. That's the part worth borrowing if you're wiring a reviewer into your own coding agent.
github.com/coderabbitai/skil…
11
Databricks' ZeroOps beta can share pipeline findings with teammates who can't open the source tables. Those findings can include sample rows and code. Before sharing an agent's report, check who can see the report as well as who can see the source.
docs.databricks.com/gcp/en/z…
10
Before making an inbox agent faster, decide which emails deserve a run. Glean now starts Outlook agents as matching mail arrives instead of polling. For invoice triage, filter by folder or sender and keep changes to the books for human approval.
docs.glean.com/release-notes…
10
Buying software and deploying it are two different decisions. AWS Marketplace’s agent can help with both. For supported deployments, it shows a plan and estimated monthly cost. That’s worth a second look even after the purchase is approved.
aws.amazon.com/about-aws/wha…
5
Start with one app, not the whole desktop. Claude’s computer-use beta is rolling out to Team and Enterprise, off by default. Workspace owners can set an app allowlist, and users still have to approve each app on it.
support.claude.com/en/articl…
2
ChatGPT dots are rolling out to eligible US Business Standard users and some Plus accounts. A useful first job: watch a supplier’s public status page and tell you only when service is back. That’s a quick way to see if you even need a dashboard.
help.openai.com/en/articles/…
25
Stripe’s new benchmark has a useful reality check: with Stripe’s plugin, Codex with GPT-6.1 Sol scored 89.6% on average, but met every required criterion in only 50.7% of runs. If AI builds your checkout, test a failed payment as well as a successful one.
stripe.dev/blog/stripe-integ…
1
20
How much work is left after a sales call? Salesforce in Copilot (public preview) proposes CRM updates from the conversation and asks before saving them. The test: how many suggestions can a rep approve without rewriting?
news.microsoft.com/source/20…
1
16
Parallel agents are useful until they start work in the wrong order. Vercel's experimental Board lets agents claim separate migration tasks, then opens each review task only after its fix is complete.
vercel.com/blog/introducing-…
19
Google says it gets 15x as many duplicate security reports as it did in 2023, and the average report is 3x as long.
If you're using AI for code review, count the distinct issues your team confirms and fixes. Report volume tells you very little.
blog.google/security/expandi…
21
A support agent can look up an order without having permission to refund it. AWS's AgentCore harness now lets you check the refund amount before the tool runs. That same check applies when the agent delegates the work.
aws.amazon.com/blogs/machine…
4
Databricks' Genie Code CLI is out of beta. A data engineer can bring a local CSV into a governed table from the terminal, while admins manage one AI budget across the CLI and workspace. Handy when teams work across both.
docs.databricks.com/aws/en/r…
18
Want an AI agent to handle appointment bookings without moving your whole contact centre?
Genesys now sells its agents separately. You can start with one type of request and keep your phone numbers, carrier and existing platform.
genesys.com/blog/post/a-fast…
28
Keep the parts of your stack that already work.
GeneXus 18 now lets an AI agent propose changes for a developer to review. Once approved, the existing generator turns those changes into code. A useful pattern for adding AI to older systems.
genexus.com/en/news/read-new…
8
LangChain says over 70% of runs on Managed Deep Agents are scheduled. Once agents start checking in on their own, it's worth tracking how often someone actually acts on a check-in. That's a better signal than how many times the agent ran.
langchain.com/blog/what-weve…
10
When an agent fails, it helps to know whether it picked the wrong tool or the tool returned bad data. Oumi's Bring Your Own Harness lets teams evaluate tool use and task completion without replacing the workflow they already run.
oumi.ai/press-release/agent-…
5
GitHub Copilot Evals is in public preview. Use /eval to test skills, custom agents and MCP tools, inspect failures, and compare behaviour before and after changes. Available in VS Code, Copilot CLI and the Copilot app.
github.blog/changelog/2026-1…
23



