In-depth analysis · About 4 min read
Foundry's two-agent changelog guide separates accepted jobs, checked reports and delivered messages
A Microsoft workflow example explains source-scoped collection, authenticated handoff and delivery receipts, without turning a schedule into an approval system.
- Published here
- Source published
- Source checked
Microsoft's Foundry September roundup, visibly dated October 15, contains a detailed two-agent developer-changelog example. This analysis focuses on that workflow rather than announcing September features again as new releases. Its key distinction is operational: accepting a job, finishing a source-checked report and delivering it to a recipient are separate states, each requiring evidence.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The producer uses GitHub Copilot SDK for scoped public-source collection; the receiver uses Microsoft Agent Framework to compose and deliver the checked report.
- Routines and Tool Search are GA, but Insights is public preview; a manual routine dispatch does not prove a calendar-triggered run.
A guide published later is not a new launch of every component
The page carries October 15 as its visible publication date even though it reviews September. Routines and Tool Search had been announced GA on September 24; Insights remains public preview. The workflow is useful as a technical guide because it explains how those components fit together, not because every component has just launched. Its Monday schedule is enabled in the described example, but the completed demonstration used manual dispatch. The source explicitly says that this is not proof of a calendar-triggered Monday run. Keeping those dates and evidence types separate avoids manufacturing a fresh release or a successful scheduled execution.
The collector has a bounded evidence task
The producer, built with GitHub Copilot SDK, collects scoped public repository changes and saves evidence and coverage gaps. The example exposes seven GitHub read operations through Tool Search, with application-level restrictions on repositories, paths and methods. A published release, merged pull request and developing branch need different evidence. Discovering a tool does not authorize comments, writes or sample execution. Hosted-agent Toolboxes remain GA, while the prompt-agent expansion is preview. AISubHub analysis: the useful first test is one source-backed note from one approved repository, not silently broadening collection whenever a required read tool is missing.
The receiver owns the handoff and channel contract
Microsoft Agent Framework powers the receiver. It adopts the checked report through authenticated Invocations and delivers through native Activity to Teams. The example separates these external contracts from model inference: Model Router supplies inference, not workflow orchestration. The receiver has its own saved report and delivery record rather than assuming access to the producer's private storage. A fixed, approved destination belongs to application code, not to the model's choice. This architecture does not imply that any operator has already installed a Teams app, registered its conversation or obtained permission to send.
Three receipts protect against false success
The application persists a job before accepting it, exposes status and result, and records delivery separately. A finished report is not a delivered message. In the reported October 13 demonstration, the producer adopted the receiver's actual message receipt without sending again. An uncertain acknowledgment remains something to investigate rather than a reason for blind replay. Saved report navigation does not trigger another collection or model call, and the example gives report and baseline records a 90-day lifecycle. That duration describes this example; it is not a general retention promise for all Foundry customer data.
Inspect behavior before changing the automation
Insights in public preview groups recurring trace behavior into findings and proposed fixes. The example includes repeated retrieval, retries after rate limits and an earlier uncertain delivery receipt. The suggested loop is to inspect a linked trace and test the proposed change on the same input before applying it. Model Router's preview session affinity also does not guarantee the same model or a cache hit. AISubHub analysis: source quality, identity, durable state and human authorization remain application responsibilities. Begin with a no-send run and verify its evidence; account for cloud and monitoring costs before enabling schedules. These are evaluation lessons from the guide, not integrations or savings we have measured on AISubHub.
Why it matters
AISubHub analysis: this design is useful for evaluating automated news and change reports because it makes the evidence handoff and delivery outcome explicit. The relevant test is whether one note traces to its source and one authorized destination, not how many jobs were accepted.
Limits and availability
This is analysis of Microsoft's published example, not a system we deployed or a message we sent. The page summarizes older September releases and describes an October 13 demonstration. Cloud inference, active compute, tracing and Insights can incur costs; no free personal entitlement follows.
Sources and evidence
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.