Foundry Hosted Agents - Tools & Toolbox
In Part 1 we deployed a hosted agent. In Part 2 we explored sessions, state, and the sandbox. In Part 3 we gave the agent an identity and a safe way to handle secrets.
F5 - Squashing Bugs
A technology blog focusing on random stuff
Latest posts
Foundry Hosted Agents - Agent Identity, Security & Secrets
In Part 1 we deployed an agent and in Part 2 we explored the sandbox — sessions, persistence, and isolation. But our agents so far have only done one thing: call a Foundry model. What if your agent needs to read files from Azure Storage, query a Cosmos DB database, or call a third-party API that requires an API key? How does authentication work when your code runs inside an isolated microvm…
Foundry Hosted Agents - Sessions, State & the Sandbox
In Part 1 we deployed a hello-world agent and had a conversation with it. It replied, remembered context across turns, and we moved on. But what actually happened behind the scenes? Where did the conversation history live? What if the agent needed to save a file — a report, a processed dataset, a set of notes — where would that go? And what happens when nobody talks to the agent for an hour?
Microsoft Foundry Hosted Agents - What, Why, Protocols & Your First Deployment
Building agentic applications is exciting — until you realize you also need to solve containerization, web servers, identity, state persistence, scaling, and observability before your agent can say “Hello.” You just want to write agent logic, not plumb infrastructure. That’s the problem that my team has solved with the launch of Microsoft Foundry Hosted Agents: you write the code, package it in a…
KV-Cache Aware Prompt Engineering - How Stable Prefixes Unlock 65% Latency Improvements
Ever wondered why some AI applications feel lightning-fast while others seem sluggish? The secret often lies in something called the KV cache – a clever optimization in transformer models that’s like having a really good memory for previous conversations.
Patterns for Vision on the Edge - Part 2 - Handling Multiple Consumers
This post is the second part in a 2 part blog series on Patterns for implementing Vision on Edge. In the previous post, we discussed two of the three common challenges faced when creating a Computer Vision based solution using Python on Edge.
Patterns for Vision on the Edge - Part 1 - Concurrent processing
While developing a solution for implementing vision on the edge use case, one of the most common requirement is having the ability to ingest live video feed and process it to derive useful insights (running a ML model on the captured frame, etc.).
Optimizing network footprint using MessagePack
The most commonly used technique to send binary data from server to client is to encode it using Base64 and send it as JSON. Base64 encodes each set of three bytes into four bytes. In addition the output will be padded to always be multiple of four. So the final output would be 4/3 or 33% larger than the original data. Add overhead of JSON to it and output grows even larger. In most of the…
Measuring latency of Websocket Messages
Have you noticed that the browser does not display time it took for a websocket message to be delivered like it shows for HTTP requests? That is because, in contrast to HTTP protocol which follows request-response pattern, websocket messages do not have transaction semantics once the initial handshake is done to establish the connection. Once the websocket connection is established, messages are…
STEADY by AID with ART
This post is heavily inspired by Hints and Principles for Computer System Design by Butler Lampson and is derived from my notes when reading the paper. My goal of this article was to condense information in Butler’s paper into short actionable article that can be referenced when designing software. I strongly recommend reading the original paper.
A better way to logging in Python
Any production application will likely have some guidelines regarding how and what needs to be logged in your application. More often these guidelines stem from common industry patterns such as “log all exceptions”. However implementing these guidelines are left to individual developers and leads to same set of logging statements repeated throughout the codebase. For instance to log all exceptions…
Azure Devops Pull Request Extension for VS Code
Marketplace: https://marketplace.visualstudio.com/items?itemName=ankitbko.vscode-pull-request-azdo.