Agent Finder and ARD are generally available in GitHub Copilot app. This guide covers task routing, agent approval, telemetry, and review controls.
What Agent Finder and ARD Add to the Copilot App
Agent Finder and ARD move the GitHub Copilot app from a single assistant toward a managed set of agents that teams can route work to and govern centrally. Agent Finder is the discovery and routing layer: it helps a request reach the agent best suited to handle it, rather than forcing every task through one generic path. ARD supplies the governance controls around that routing — approval, telemetry, and review — so an organization can decide which agents are allowed to act and under what conditions.
The practical shift is that agent behavior becomes something you configure and audit, not just something you invoke. That matters most for teams that need a defensible record of what an automated agent did, why it was permitted to do it, and who signed off.
Task Routing and Agent Approval
Task routing decides which agent receives a given request. When you set this up, think in terms of the kinds of work you actually see: code generation, refactoring, review, documentation, or operational queries. Map each category to an agent whose scope fits, and keep the routing rules narrow enough that a request lands somewhere predictable instead of defaulting broadly.
Agent approval is the gate that sits in front of routing. Rather than trusting every available agent by default, treat approval as an allowlist: an agent should be usable only after someone with authority has reviewed its scope and enabled it for the team.
- Approve agents against a specific scope of work, not a general "trusted" flag.
- Route ambiguous or high-impact tasks to agents that require an extra confirmation step.
- Re-review approvals when an agent's capabilities or permissions change.
- Keep an explicit fallback so an unroutable request fails visibly instead of silently.
Telemetry and Review Controls
Telemetry is what makes governance real instead of aspirational. Capture which agent handled each task, what routing decision led there, and the outcome. The goal is a trail you can query later — to spot an agent being used outside its intended scope, to see where routing sends the most volume, and to justify decisions during an audit or incident review.
Review controls close the loop by putting a human checkpoint where the risk warrants one. Not every task needs manual review, so tune the controls by impact: low-risk, easily reversible actions can flow through, while changes that touch production paths, permissions, or sensitive code should pause for a person. Use telemetry to inform where those checkpoints belong rather than applying one blanket rule everywhere.
Rolling It Out
Start narrow. Enable Agent Finder and ARD for a small set of well-understood task types, approve only the agents you need, and turn on telemetry from day one so you have data before you widen access. Watch how routing behaves against real requests, then adjust approval scopes and review thresholds based on what the telemetry shows rather than on assumptions.
Because these features are now generally available, they are meant to be treated as standing infrastructure, not an experiment. Document your routing rules and approval decisions where the whole team can see them, and revisit the configuration on a regular cadence so it keeps matching how your team actually works.