MS Principal Cloud Advocate. Connecting people, places & things to enable everyone. #Microsoft #AZOPS #ITPRO #IOT #Security #AI #MSIgnite #MSBuild #EDU
Toronto, Ontario
- Tweets 15,240
- Following 1,236
- Followers 5,730
- Likes 11,525
The next AI agent fix may be hiding in a pattern, not one trace.
Foundry can surface recurring behaviors and optimize instructions, skills, tools, and model choices.
Join the AMA: aka.ms/FoundryFriday-Optimiz…
ALT Microsoft Foundry Friday AMA
95
MCP's biggest win is not smarter AI. It is fewer one-off integrations.
3 AI apps x 3 systems = 9 connections to maintain.
MCP replaces that tangle with a shared protocol.
Which system would you expose first through MCP?
techcommunity.microsoft.com/… #MCP
ALT MCP Servers Explained
147
A prompt that says "ask first" is not an approval control.
@lee_stott shares if an agent can write to a real system, the runtime must pause, collect a decision, then resume or reject that exact operation.
Where is your approval gate enforced today?
techcommunity.microsoft.com/…
1
2
170
Permission prompts are not a security boundary. KARS puts control outside the agent: no direct egress, separate router-held credentials, policy and audit.
The key shift is architectural containment.
Where is your trust boundary?
techcommunity.microsoft.com/…
3
137
The best-looking image model can still be the wrong production model.
Gallery samples show the ceiling. Production depends on the floor: prompt adherence, spatial reasoning, safety, bias, latency.
What failure would disqualify it?
techcommunity.microsoft.com/…
ALT Evaluating Text-to-Image Models
1
98
Smarter agents won't finish workflows if the last step lives in a browser.
Microsoft's Playwright Workspaces remote MCP server adds 22 browser tools and an observe-act-verify loop.
Where does your automation still hit a browser-only wall?
techcommunity.microsoft.com/…
2
1
4
296
A correct answer can still invalidate an AI eval.
If the agent found the answer in local source, caches, tools, or services, you measured retrieval, not model knowledge.
Inspect surprising passes.
What route would break your eval?
developer.microsoft.com/blog…
1
1
1
194
Publishing a Copilot Studio app doesn't expand a user's data access. The app acts on behalf of the signed-in user, so permissions still apply. Admin focus shifts to inventory, credit caps, and build paths. Which control comes first?
techcommunity.microsoft.com/…
1
2
164
Dashboards are becoming executable context for AI agents.
One Grafana dashboard gave an agent the queries, scopes and pitfalls to diagnose two "long" sessions as different problems.
What operational map does your agent already have?
techcommunity.microsoft.com/…
184
AI should not decide who owns a support case. Let the model propose. Let the flow validate. Let a person approve. In this lab, Outlook only creates an editable draft after approval. Where would you put the human checkpoint?
techcommunity.microsoft.com/…
3
178
AI coding gets safer when verification gets easier.
Diff: what changed?
Terminal: does it run?
Browser: does it actually work?
That is the loop before accepting agent-made code.
Which check do you skip most often?
github.blog/ai-and-ml/github…
2
183
GPT-Live-1 can listen and speak at once; GPT-image-2.5 can refine visuals across turns. That's a shift from prompt-response to live collaboration. What would you redesign first?
techcommunity.microsoft.com/…
1
1
195
No code change does not mean no AI behavior change.
Microsoft Foundry's Model Router refreshed its pool with GPT-5.6 variants and Claude Opus 4.8 while retiring older models.
Re-test quality, latency, cost. What would you benchmark first?
Read: devblogs.microsoft.com/found…
1
183
Paying more for a model does not guarantee better work.
JDK migration test: Sonnet scored 97% at $12.71/run. Astra scored 93% at $67.33/run.
Same skill. Higher price. Lower score.
What gain would justify the premium?
developer.microsoft.com/blog…
#AIAgents #LLMOps
ALT Claude Sonnet vs GPT-6 Astra Model comparison
1
1
197
Choosing the model is not the same as being ready to ship.
Microsoft's Sparkles demo goes beyond prompts: tools, traces, and evaluation.
A prototype can answer. A production agent has to earn confidence.
What's your bottleneck?
techcommunity.microsoft.com/…
1
1
120
In scientific AI, the best model may matter less than knowing when to switch.
CLIO treats failed paths as evidence and changes course, including switching models when needed.
Which scientific workflow would you test this on first?
techcommunity.microsoft.com/…
126
Passing tests does not make an AI agent safe. Bound its authority, not just its output.
Read: techcommunity.microsoft.com/…
151
Microsoft Copilot's August update is about control: effort levels, credit visibility, brand rules, and change history.
The AI question is shifting from "Can it do the work?" to "Can we govern the work?"
Which control matters most?
Read more: techcommunity.microsoft.com/…
2
185
AI training works better when it does not pull you away from the work.
Microsoft's Learning Agent can now surface AI Skills Navigator modules, videos, learning paths, and credentials inside Copilot.
Which skill gap would you test first?
techcommunity.microsoft.com/…
133
The benchmark can be wrong even when every number is right.
A 25.8s p95 at the deployed-agent boundary cannot be judged against a <=30s Copilot Studio client gate.
Before you rank retrieval patterns, verify the measurement boundary.
Read: techcommunity.microsoft.com/…
128