|
|
|
Every thread is a full cloud session on its own branch, which is why per-project usage tracking shipped in the same release.
|
|
Four vendors shipped agent orchestration this week, and each one shipped a permission checkpoint next to it. Orchestration is moving off your laptop and into vendor infrastructure, so the line worth hunting for in every release note is where a human still has to sign.
|
|
|
|
|
GitLab 19.4 opens the DevOps loop to MCP clients - with writes behind an approval
GitLab 19.4 landed September 17 and opened the DevOps loop to outside MCP clients including Claude Code, Cursor, and VS Code: pipelines, merge requests, issues and epics, repo commits, vulnerability triage. Read-only tools default to Always Allow, write and delete tools default to Always Ask - set per tool in group and project GitLab Duo settings. Free, Premium, and Ultimate all get it.
Why it matters. An agent that reads pipeline logs unprompted but cannot merge without a human is the shape a security team signs off on. See which tools default to Always Ask →
|
n8n 2.40 moves MCP tool calls onto queue workers and adds Force Tool Call
MCP toolkit members now execute on n8n's distributed worker processes, lifting the single-process ceiling that capped MCP throughput on self-hosted stacks in queue mode. Same September 15 release: Confluence is out from behind its feature flag as an AI Agent tool, and the AI Agent node (v3) gained Force Tool Call on First Iteration for smaller models such as Mistral-Small and local Ollama that answer in prose and ignore wired tools.
Why it matters. Force-tool-call is one toggle that turns a flaky small-model agent into a predictable one, and it ships off by default. Queue mode means running worker processes somewhere - Cloudways if you want managed servers, or Make if you would rather not run the substrate at all. Read the 2.40 breakdown →
|
Power Apps canvas authoring hits GA, MCP server opens in preview
The canvas authoring agent plugin is generally available: an agent creates, inspects, and updates canvas apps in a live session while the maker decides what gets handed off. The model app-builder skill also reached GA and runs through GitHub Copilot CLI or Claude Code. A separate Power Apps MCP server is in public preview: agents pull fields from shared mailboxes and SharePoint and create records, with a maker approving each pending change before it commits. Preview scope is data entry only and US early-release environments first.
Why it matters. GA kills the no-SLA objection that kept agentic authoring out of production in Microsoft shops. See what GA actually covers →
|
OpenAI Sponsored Agents open a conversation instead of a landing page
Sponsored Agents went into test with select US advertisers on September 16. Clicking an ad in ChatGPT opens a clearly labeled conversation with a business-sponsored agent, kept separate from ChatGPT's own answers and from the thread the user was already in. HubSpot is the first CRM partner, Shopify the first ecommerce partner; US Shopify merchants can install the ChatGPT Ads app now, other markets on September 23.
Why it matters. The conversation becomes the conversion event, which quietly invalidates the landing-page funnel you tune every quarter. HubSpot and Shopify sellers in the US can run this today. Everyone else should be cleaning up product data, because that is what the agent answers from. Read what this changes for attribution →
|
|
•
|
Snap shipped its assistant outside Snapchat. Specs Intelligence launched September 16 as its own iOS app, reading calendar, inbox, and travel context from connected Google accounts, with stated commitments not to train on that data or target ads against it. US limited preview, Google accounts only, read-only. Snap Specs Intelligence details →
|
|
•
|
OpenAI matched Anthropic on embedded evaluators. Both labs committed on September 16 to giving outside evaluators ongoing, employee-level access to training pipelines, logs, and intermediate checkpoints rather than finished models only, with METR and Redwood Research named as examples. Each lab still picks its reviewers and controls publication. Embedded evaluators details →
|
OpenAI Agents API vs Anthropic Managed Agents (September 2026)
Priced on the same synthetic job (1,000 sessions, 200,000 tokens each, 20 web searches, ten minutes of runtime), the two land about 7 percent apart - roughly $1,200 against $1,293 at September 17 published rates. Price is a wash, so two other rows decide it: OpenAI is US-only with no Zero Data Retention at launch, out for regulated or non-US data, and Anthropic caps spend per session where OpenAI caps only the org. See the decision table and the cost math →
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|