The agent stack is hitting a very normal software problem: too many tools.
GitHub’s June 17 Copilot updates are useful because they do not pretend the answer is to shove everything into the context window and hope the model figures it out. Agent Finder gives Copilot a way to discover relevant MCP servers, skills, canvases, agents, and tools from a registry. A user describes the task, the system searches available resources, and Copilot can pull in ranked matches on demand.
That sounds small. It is not small.
Discovery becomes infrastructure
The first wave of agents was about capability. Can the model code, click, search, read logs, call APIs, and produce a useful result?
The second wave is about placement.
Which tool should the agent use for this task? Which registry is allowed? Which internal resources exist? Which ones are approved? Which ones should stay out of reach? Which tool is relevant now, and which one is just context bloat pretending to be optionality?
GitHub says Agent Finder uses the open Agentic Resource Discovery specification and can point at a public catalog or a private enterprise registry. It also says the feature does not silently install or connect tools.
That boundary matters because discovery without governance becomes chaos, while governance without discovery becomes a locked cabinet nobody uses.
Enterprises will not accept yolo mode
The same batch of updates includes enterprise-managed settings to disable automatic permission bypass in Copilot CLI and VS Code. In plain language: an admin can prevent users from turning on the mode where the agent skips approval prompts.
That is the unglamorous detail that decides adoption.
Teams do not only worry that an agent will make a bad suggestion. They worry that it will run a command, touch a file, call a tool, or approve a change before anyone has a chance to understand the blast radius.
So the future agent interface needs two opposite things at once. It needs to find the right capability quickly, and it also needs to prove that capability is allowed, scoped, inspectable, and not quietly wired into the workflow behind everyone’s back.
Model choice is becoming routing
GitHub also made Copilot’s auto model selection generally available. Auto mode routes requests to models based on task complexity, availability, plan, and policy.
That fits the same pattern.
The user does not want to become a full-time dispatcher. The enterprise does not want unmanaged chaos. The agent runtime needs to choose resources, models, and tools dynamically while respecting policy.
That is where agent products are going: less manual wiring, more registries, routing, permissions, and auditability.
The assistant is no longer just a box where you type instructions. It is becoming a controlled runtime that discovers what the work needs, loads only what is allowed, and leaves enough trail for humans to stay responsible.
That is less flashy than another demo, and much closer to real software.
Sources: GitHub Agent Finder, GitHub permission controls, GitHub auto mode