Inside the Microsoft Copilot Harness: Customizing AI Agents
Andrea Griffiths and Patrick Nikoletich discuss how teams can adapt the Copilot harness to different agent deployments, with demonstrations of runtime isolation and remote CLI access.
Customizing the harness for agent workloads
The discussion moves from coding agents and personal assistants to fully autonomous agents, including non-coding scenarios and autonomous software factories. Patrick covers customization across three use cases and the design of a secure runtime that supports flexible agent deployments.
Worker pools, credentials, and remote access
Patrick demonstrates a Dockerized agent worker pool and explains the security architecture around it:
- Isolated containers provide environments for agent workers.
- Credential-safe inference gateways handle inference access without exposing credentials to the workers.
- Agent Host Protocol (AHP) tunnels CLI sessions into isolated containers.
The accompanying demo repository includes a harness builder and a secure Aspire agent service.
Planned changes in Copilot Harness 2.0
The episode also discusses upcoming work on Copilot Harness 2.0, including a Rust rewrite and direct Azure tunneling, alongside open-source access. These are presented as planned changes rather than capabilities demonstrated in the current worker-pool setup.
Episode chapters
- 00:00 — Introduction
- 01:05 — From coding agents and personal assistants to fully autonomous agents
- 02:55 — Designing a secure runtime for flexible agent deployments
- 04:52 — Customizing the harness for three different use cases
- 08:11 — Dockerized worker pools and security architecture
- 10:07 — Agent Host Protocol and CLI tunnels
- 12:20 — Direct Azure tunneling and open-source access
- 13:27 — What is coming in Copilot Harness 2.0