Microsoft has opened the Copilot harness in Azure Foundry Agent Service for public preview and announced general availability of long‑running resilience for hosted agents. The harness supplies a managed agent loop, sandbox, and declarative definition model, allowing teams to build reasoning agents without constructing the underlying orchestration stack.
Key Changes in the Copilot Harness
The preview bundles four reusable components:
- Reusable skills – packaged instructions, scripts, and resources that an agent can invoke across multiple tasks.
- Connected tools – integration points for Foundry and Model Context Protocol (MCP) tools, each requiring explicit authentication and approval configuration.
- Execution and delegation – built‑in support for file and code handling, plus the ability to offload bounded work to sub‑agents while the primary agent coordinates results.
- Source‑controlled definitions – agent definitions can be authored in Foundry or stored alongside application code for CI/CD workflows.
Developers select a supported Foundry model and the harness automatically manages working context, step coordination, and policy limits.
Architectural Impact
From an architecture perspective, the harness abstracts the agent loop and sandbox, shifting responsibility for runtime orchestration from the application team to the platform. This reduces the surface area of custom code that must handle model‑tool coordination, context persistence, and error handling for multi‑step workflows.
Because tool connections are declared with explicit auth and approval settings, the platform can enforce isolation between the agent and external services. The ability to delegate to sub‑agents introduces a hierarchical execution model that can be leveraged for bounded, sandboxed tasks such as code interpretation or file manipulation.
Operational and Security Considerations
Operationally, the managed loop means fewer moving parts to monitor. Long‑running resilience, now generally available, adds automatic retry and state‑preservation for agents that exceed typical request timeouts, improving reliability for complex workflows.
Security implications stem from the explicit authentication required for each connected tool. Engineers must configure credentials, approvals, and policy boundaries for each tool integration, ensuring that the agent cannot exceed its authorized scope. The sandboxed execution environment further isolates agent code from the host infrastructure.
Source‑controlled definitions enable versioned agent configurations, allowing standard DevOps practices—code review, automated testing, and pipeline deployment—to be applied to agent logic. The example below shows a minimal Azure Developer CLI (azd) manifest that defines a reusable skill and an agent that uses web search and a code interpreter:
name: research-brief
services:
ai-project:
host: azure.ai.project
endpoint: ${FOUNDRY_PROJECT_ENDPOINT}
research-brief-skill:
host: azure.ai.skill
uses:
- ai-project
archive: ./skills/research-brief
research-brief:
host: azure.ai.agent
project: .
uses:
- ai-project
- research-brief-skill
kind: prompt
harness: type: microsoft_copilot_preview
model: ${MODEL_DEPLOYMENT_NAME}
skills:
- research-brief-skill
instructions: |
Follow research-brief to answer the user’s question.
tools:
- type: web_search
name: public-research
- type: code_interpreter
name: research-analysis
container: type: auto
infra: provider: microsoft.foundry
Deploying and invoking the agent is reduced to two azd commands, illustrating the streamlined workflow.
Related CloudNinjas coverage: Azure.
What This Means For Practitioners
Engineers should evaluate the harness as a way to offload orchestration, focus on domain‑specific knowledge, and adopt the built‑in resilience for long‑running tasks. Key actions include:
- Map existing agent pipelines to the four harness building blocks and identify which components can be replaced by managed equivalents.
- Review authentication and approval settings for each connected tool to ensure they meet organizational policy.
- Integrate agent definition files into version control and CI/CD pipelines to benefit from automated testing and rollout.
- Monitor the preview rollout for GA updates, additional model support, and expanded policy controls.
By leveraging the Copilot harness, teams can accelerate agent development while maintaining control over execution context, tool access, and operational reliability.
