07Video
Tisha Talks Tech
On YouTube I break down how agentic systems work, walk through real engineering problems, and share what I've picked up building at Microsoft.
Tisha Talks Tech
AI agents, engineering deep dives, and career notes.
Speaking at AI Engineer World's Fair 2026
Tisha Chawla
Software engineer at Microsoft. I work on making AI agents reliable in the real world, the part that decides whether they hold up outside a demo.
01About
I build data platform infrastructure at Microsoft, where my work sits where large-scale analytics meets agentic AI. Most of what I actually do is make agents boringly reliable: durable state so they don't forget, deterministic execution so a failure can be reproduced, and self-healing workflows that recover without someone getting paged at 2am.
I treat AI as something to architect, not just consume. The interesting problems usually aren't in the model, they're in the orchestration around it: how coding agents, reliability agents, and spec-driven workflows actually hold together once real traffic hits them, rather than in a tidy demo.
Requirements are no longer static documents. They are the first architectural control surface, where systems inherit intent before implementation begins.
From my LinkedIn series on the agentic SDLCOn the research side, I publish peer-reviewed applied-ML work with 150+ citations. And I do most of this in the open, through my blog, talks, and a running LinkedIn series.
02Focus
Agents across the lifecycle
Reliability isn't one feature you bolt on at the end. It's something you engineer at every stage, from the moment a requirement is written to the moment an agent quietly recovers from an incident on its own.
01
Requirements
Spec-driven dev
02
Build
SWE agents
03
Test
Testing agents
04
Operate
SRE agents
05
Scale
Cost & reliability
Spec-driven development
Requirements as the first control surface. I work with Spec Kit across greenfield and brownfield so intent flows cleanly into what gets built.
SWE agents
The Azure DevOps SWE Agent and custom agents: architecture, MCP integration, and automating pull request creation and implementation.
Testing agents
Playwright Planner, Generator, and Healer agents that write and self-heal test cases automatically, in VS Code and Azure Foundry.
SRE & reliability agents
The Azure SRE Agent: incident automation, ICM and DevOps integration, and sub-agent builder workflows that keep things dependable.
AI-first repositories
Making repos genuinely agent-ready through structure, docs, and context engineering, including Microsoft's first-party Agency platform.
Agent economics
FinOps and token optimization for non-deterministic agents, so the bill stays visible and under control as systems scale.
03Speaking
Talks
I share what I'm building with engineering audiences inside and outside Microsoft.
FinOps for AI Agents: Who Spent All the Tokens?
A practical look at cost management and token optimization for non-deterministic agents: where the spend actually goes, how to make it visible, and how to design agents that stay economical as they scale.
Conference details →2026 · Online
Your Agent Failed in Prod. Good Luck Reproducing It.
AI Engineer World's Fair 2026, online track
20 May 2026
Spec-Driven Development with Spec Kit
Microsoft APAC AI Forum
AI Driven Engineering series, Microsoft
Sessions I presented in Microsoft's internal AI Driven Engineering (AIDE) series, covering agents across the development lifecycle.
Azure DevOps SWE Agent
Architecture, custom agents, MCP server integration, features, and prompting best practices.
Azure SRE Agent
Onboarding, incident automation, ICM and DevOps integration, and sub-agent builder workflows.
Spec Kit, SWE Agent & Playwright Testing Agents
Brownfield migration with Spec Kit, automated PR creation, and Playwright Planner, Generator, and Healer agents.
The Agency Platform & AI-First Repos
Microsoft's first-party agent platform and patterns for making repositories truly AI-first.
GitHub Copilot CLI & SDK with Liquid MCP
Copilot SDK and CLI, MCP server integration, workflow handling, and compliance guidance with Liquid MCP.
04What I believe
A few convictions
These aren't slogans. They're the lessons that keep showing up in my work, and they shape how I build.
01
Reliability is the product.
With agents, the model is the easy part. Whether it holds up under real traffic is what people actually feel and remember.
02
If you can't reproduce it, you can't fix it.
Non-determinism is the enemy of debugging. I design for reproducible runs so a prod failure is a lead, not a mystery.
03
Boring infrastructure wins.
Durable state and self-healing workflows are unglamorous, but they're what separate a demo from a system you can trust.
04
Research and engineering belong together.
The best systems come from both: rigour from research, pragmatism from shipping. I try to live in that overlap.
05How I work
I build in the open
The fastest way I've found to get good at something this new is to work on hard problems and then actually show the work, what shipped, what broke, and the messy reasoning in between. So I write the long post, give the talk, and publish the thing that isn't perfect yet.
Build at the frontier · problems without settled answers
Learn in the open · the process, not the highlight reel
High signal · depth and evidence over buzzwords
06Writing
Why agents break, and how to fix it
My blog goes deep on the hard parts of agent engineering: memory, reproducibility, and structure. A few recent ones:
08Contact
Let's talk
Speaking invitations, collaborations, or just to say hi.
09Building
The Agent Plane
An open-source effort focused on the infrastructure that makes AI agents reliable in the real world.
The Agent Plane
Open infrastructure for running AI agents reliably, from durable state to reproducible execution. A work in progress, built in the open.
View on GitHub →