Your own note near the end, that the AI Orchestration setting isn't present in your environment, is very likely the actual root cause here, more so than anything on the Databricks/MCP side.
The official Databricks integration guide for this exact setup is explicit that this isn't optional: "In the upper-right, click Settings. Under the Orchestration section, click Yes to use AI orchestration for your agent's responses." Without that toggle enabled, my understanding is Copilot Studio falls back to its classic topic/trigger-based flow rather than an LLM-driven planner that can autonomously decide mid-conversation to call an arbitrary tool. That would explain your exact symptom precisely: the MCP tools are correctly registered and discoverable (which is a static/config-time thing), but nothing at runtime is actually deciding to invoke them, because the component that does that deciding is the thing that's missing from your setup.
I don't have a confirmed reason why that Orchestration option itself would be absent from your environment specifically, since Databricks' docs describe it as a standard step rather than something gated. Worth checking whether the agent was created through the same flow the docs assume, since Copilot Studio has more than one agent-creation path and they don't all expose the same Settings panel, and if it's genuinely missing and not just relocated, that's likely a Power Platform tenant/feature-rollout question rather than a Databricks-side one, so Microsoft support or your Power Platform admin center may be able to confirm whether AI orchestration is enabled for your tenant/region.
Separately, and worth doing regardless once orchestration is on: Microsoft's own troubleshooting guidance for "tool discovered but not invoked" also points at the tool's description text specifically, vague descriptions can stop the orchestrator from ever selecting a tool even when orchestration is active. So once you have the setting available, it's worth opening query_space_<agent_id> in the Build tab and giving it an explicit description naming CMGs, generation, demand, policies, CAPEX, IFOR, etc, rather than relying only on your agent-level instructions to carry that context.
Two things point at the same root cause. First, the exact failure mode you're seeing, discovered but never invoked, is Microsoft's own documented troubleshooting entry for agent tools: "Tool not being invoked: Review the tool's description to make sure it clearly describes when the tool should be used. Vague descriptions can prevent the orchestration runtime from selecting the tool." That's from the Copilot Studio docs on managing tools in an agent. Second, your agent's error text, "I only have local tools such as bash, grep, file utilities, etc.", is the signature phrasing of an agent running on the GitHub Copilot harness inside Copilot Studio, a different underlying runtime from the classic Power Virtual Agents-style Copilot engine. On that harness, tool selection is planner-driven off the tool's own description, separate from your agent's system instructions, so instructing the agent to "use Genie for CMGs, generation, demand, policies, CAPEX, IFOR" in its own prompt doesn't help if the query_space_<agent_id> tool itself just has Databricks' generic auto-generated description with none of that domain vocabulary in it.
The fix per the docs: Build tab, Tools panel, select the MCP server, open query_space_<agent_id>, and edit its name/description directly so it explicitly states what it's for, something like "Query the PLANI GENIE space for questions about CMGs, generation, demand, policies, CAPEX, IFOR, or any Databricks table this space covers." Do the same for poll_response_<agent_id> if it has a similarly generic description. That's a per-tool edit in the Tool details dialog, separate from your agent's own instructions field, and it's specifically called out as the fix for "discovered but not invoked."
Worth checking which harness your agent is actually running on if you haven't already, since the fix path differs slightly between the GitHub Copilot harness and the classic Copilot Studio engine, but the underlying cause, orchestrator not being told when to use the tool, is the same either way.
