Scaling & Coordinating AI Agents for Development - Tamir Dresher

Over the past year, Tamir have been working daily with coding agents, not as a single assistant, but as a coordinated team of AI “developers” running in parallel. Some design, some implement, some review, some break things, and some fix them. The challenge quickly stopped being what can an agent do? and became how do you run many of them effectively at the same time?
In this session, Tamir shares the concrete practices, workflows, and mental models he developed to operate multiple coding agents like a real development team.
We’ll walk through how he splits work between agents, how he writes tasks so agents don’t step on each other, how he manages parallel branches and workspaces, how he synchronizes progress, and how he prevents chaos, wasted cycles, and hallucinated architecture.

Rather than theory, this talk is about what actually works in day-to-day engineering:
how to turn agents into specialized teammates, how to coordinate them across features, refactors, and investigations, and how to scale your personal output without losing technical control

Transcript

1. Scaling & Coordinating AI Agents for Development Tamir Dresher Principal Engineer @ Microsoft www.tamirdresher.com https://www.linkedin.com/in/tamirdresher/ https://github.com/tamirdresher/worktrees-example
2. Who am I? My Books: .NET Bond Leadership team Ex-Microsoft MVP Tamir Dresher Currently: • Principal Engineer @ Microsoft Security • Software Engineering Lecturer @ Ruppin Academic Center Previously: • Head of Architecture @ Payoneer • Chief Architect @ Clarizen
3. My favorite Coding Agent
4. My favorite Agent – Roo code
5. Working with tools - approvals
6. Modes & instructions
7. Modes & instructions
8. Parallelizing work
9. Git worktrees
10. Git worktrees
11. Automating the worktrees
12. Automating the worktrees
13. Automating the worktrees
14. Starting the plan while VS Code starts  Github Copilot CLI
15. Starting the plan while VS Code starts  Github Copilot CLI
16. Github copilot worktree support
17. Giving your agent eyes
18. Giving your agent eyes { "mcpServers": { "playwright": { "command": "npx", "args": [ "@playwright/mcp@latest", "--browser=msedge", "--extension" ], "alwaysAllow": [ "browser_click", "browser_navigate", "browser_snapshot", "browser_screenshot" ] } } }
19. Giving your agent eyes { "mcpServers": { "playwright": { "command": "npx", "args": [ "@playwright/mcp@latest", "--browser=msedge", "--extension" ], "alwaysAllow": [ "browser_click", "browser_navigate", "browser_snapshot", "browser_screenshot" ] } } }
20. Talking with your agent
21. Giving your agent mouth and ears  Windows logo key+H  VS Code extension – https://marketplace.visualstudio.com/items?itemName=ms-vscode.vscode-speech  VoiceMCP - https://github.com/tamirdresher/mcp-voice-assist
22. Giving your agent mouth and ears https://www.youtube.com/watch?v=wP6m5xLAbOs
23. Running agents in the background
24. Microsoft DevBox - https://devbox.microsoft.com/  a cloud-based development environment that gives you powerful dev machines in the cloud.  Work continues even when I close my laptop  I can spin up multiple machines for different projects or experiments  I can access my full dev setup from anywhere, even from my phone using Windows App  Pro-tip: Live Share
25. Github copilot background work
26. Working on distributed systems
27. Aspire - https://aspire.dev/ Opinionated, cloud ready stack for building polyglot observable, production ready, distributed applications Aspire Composition & Orchestration Components Tools • Modeling language for defining the system resources and their dependencies • A runtime to for running and connecting multi-project applications and their dependencies • Dashboard • Distributed App VS support • Service Defaults • Publishing utils • Packages of commonly used services, such as Redis or Postgres, with standardized interfaces and capabilities
28. Aspire apphost
29. Aspire apphost
30. Aspire apphost
31. Aspire apphost
32. Aspire apphost
38. Making each worktree – aspire isolated Repo worktree1 worktree2 worktree3 . ├── scripts/ │ ├── start-apphost.ps1/sh # Starts AppHost with auto-port allocation │ │ │ ├── list-apphosts.ps1/sh # Lists all running AppHost instances & PIDs │ │ │ ├── stop-apphost.ps1/sh # Gracefully stops a specific instance │ │ │ ├── kill-apphost.ps1/sh # Force kills instances (specific PID or --All) │ │ │ └── aspire-mcp-proxy.cs # MCP Server Proxy │ └── src/ └── NoteTaker.AppHost/ # The Aspire Orchestrator project
39. Making each worktree – aspire isolated Repo worktree1 worktree2 worktree3 . ├── scripts/ │ ├── start-apphost.ps1/sh # Starts AppHost with auto-port allocation │ │ │ ├── list-apphosts.ps1/sh # Lists all running AppHost instances & PIDs │ │ │ ├── stop-apphost.ps1/sh # Gracefully stops a specific instance │ │ │ ├── kill-apphost.ps1/sh # Force kills instances (specific PID or --All) │ │ │ └── aspire-mcp-proxy.cs # MCP Server Proxy │ └── src/ └── NoteTaker.AppHost/ # The Aspire Orchestrator project
40. Aspire MCP Isolation problem
41. Aspire MCP Isolation problem Solution: The aspire-mcp-proxy Fundamental Theorem of Software Engineering (FTSE) “All problems in computer science can be solved by another level of indirection... Except for the problem of too many layers of indirection.” David Wheeler
42. Aspire MCP Isolation problem Solution: The aspire-mcp-proxy
43. Aspire MCP Isolation problem - fix
44. My agent is now independent!
45. My agent is now independent!
46. My agent is now independent!
47. Distributed Testing
48. Finishing the work
49. Finishing the work
50. Finishing the work
51. Summary  Worktrees allows to parallel the work  Background agents via DevBox or Codespaces etc will multiply your capacity  Aspire gives the agent distributed systems powers  Instructions & Modes  Glues everything together
52. Thank you www.tamirdresher.com https://www.linkedin.com/in/tamirdresher/