Custom subagent model in VS Code Copilot seems to fall back to parent model

[cbdeveloper]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

Models

Body

Custom subagent model in VS Code Copilot seems to fall back to parent model

Hi, I’m trying to understand whether this is expected behavior or a bug.

I’m using custom agents in VS Code Copilot with .agent.md files.

My setup is:

• Main chat session is running on GPT-5.4
• One workflow step delegates to a custom agent
• That custom agent has model: "Claude Opus 4.6" in its .agent.md

What I expected:

• The delegated custom agent/subagent would run on Claude Opus

What I’m seeing:

• When I hover the delegated run in the UI, it still shows GPT-5.4

So I’m not sure which of these is true:

1. The custom agent model override is working, but the UI hover only shows the parent model
2. The custom agent model is not being honored and it is falling back to the parent model
3. My model string is not in the exact format VS Code expects

My main questions are:

1. Are custom .agent.md agents in VS Code supposed to be able to override the parent model when used as subagents?
2. If yes, should the hover show the subagent’s real model, or only the parent session model?
3. Does the model field need an exact qualified name like (copilot) to work properly?
4. If the model name does not resolve, does Copilot silently fall back to the parent model?

If anyone has this working, an example of the exact model: format would help a lot.

[Aninda2002]

Can .agent.md override the parent model?

Short answer:
👉 Yes, but only if the model is actually supported and correctly resolved by Copilot.

However:

Copilot does not guarantee strict enforcement of model: in subagents
In many cases, routing is advisory, not absolute
If the model isn’t available → fallback happens

Why hover still shows GPT-5.4?

This is the most confusing part.

👉 The hover UI typically shows:

The parent chat session model, not the actual execution model of subagents

So:

Seeing GPT-5.4 in hover ≠ proof it's not using Claude
It just means the UI is not granular enough

💡 Think of it like:

“Session model = GPT-5.4”
“Execution model = maybe overridden (but not visible)”

Your biggest red flag: "Claude Opus 4.6"

This is likely the real issue.

Copilot does NOT accept arbitrary human-readable model names.

It usually expects:

Fully qualified internal IDs
Provider-prefixed names
❌ Likely invalid:
model: "Claude Opus 4.6"
✅ More realistic formats:
model: anthropic/claude-opus-4.1
or (depending on setup):

model: claude-opus-4

or GitHub Copilot-specific routing labels like:

model: gpt-4o
model: o3

👉 “4.6” especially looks suspicious — Anthropic doesn’t publicly use that version label in most integrations.

What happens if model is invalid?

Yes — and this answers one of your key questions:

👉 Copilot will silently fall back to the parent model

No error
No warning
No UI indication

So your situation strongly suggests:

❗ The model string is not resolving → fallback to GPT-5.4

How to verify if override is actually working

Since UI doesn’t help, you need behavioral testing:

Try this:

Inside your subagent prompt:

Respond with exactly:
MODEL_CHECK:

Then compare:

Claude-style output (more verbose, structured, cautious)
GPT-style output (more direct, slightly different tone)
Or even:

What model are you?

(Not always reliable, but sometimes reveals routing)

Working .agent.md example

Here’s a safer template:

---

name: my-claude-agent
description: Uses Claude for deep reasoning
model: anthropic/claude-3-opus

You are a deep reasoning agent...

👉 Adjust based on what your Copilot plan actually supports.

🧩 7. Answers to your exact questions
✔️ Are subagents supposed to override model?

👉 Yes — but only if the model is valid and available

✔️ Should hover show subagent model?
No (currently)
It usually shows only the parent session model

✔️ Does model: need exact format?

👉 YES — absolutely critical

Must match provider-supported ID
Human names won’t work
✔️ Silent fallback?

👉 Yes — 100% confirmed behavior
If:

Model not found
Model not enabled
Wrong format

→ falls back silently

🎯 Final Diagnosis

Your case is almost certainly:

❌ "Claude Opus 4.6" is not a valid model identifier
→ Copilot cannot resolve it
→ silently falls back to GPT-5.4
→ hover shows GPT-5.4 (parent model)
What I’d do next

Replace with a known valid model ID, e.g.:

model: anthropic/claude-3-opus
Check:
Is Anthropic enabled in your Copilot/org?
Does your plan allow multi-model routing?
Run a behavioral test prompt

[cbdeveloper]

It's not that the hover is wrong.

If I inspect what the subagent is doing, it's pretty clear that it's indeed using the parent model.

And if you open a new chat and select the custom agent directly, it uses the correct model specified in its .agent.md definition file.

[Aryan-Gore]

what you’re seeing is usually expected—your subagent is likely falling back to the parent model, not actually running on Claude.

Here’s how it works in Visual Studio Code with GitHub Copilot:

1. Subagents don’t reliably override models (current behavior)

In most setups, .agent.md does not guarantee a model switch when invoked as a subagent.
The parent chat model (GPT-5.4 in your case) is typically used unless the platform explicitly supports multi-model orchestration.

2. UI hover is a strong signal

If hover shows GPT-5.4, that usually means the subagent actually ran on GPT-5.4, not Claude.
It’s not just a display issue in most cases.

3. Model string format matters—but support is limited

Even if you write model: "Claude Opus 4.6", it must match a supported, fully qualified model ID recognized by Copilot.
Unsupported or unknown model names are silently ignored.

4. Silent fallback does happen

Yes—if the model isn’t recognized or allowed, Copilot will quietly fall back to the parent model (no error).
What to check
Try using an officially supported model identifier (from Copilot docs, not a plain name)
Confirm your Copilot plan/org actually has access to that model
Test by switching the main chat model to Claude—if that works, subagents are not overriding
Bottom line
Model override in .agent.md for subagents is not reliably supported
Your case is most likely fallback, not UI inconsistency

If you want, I can help you test a minimal .agent.md that confirms whether override is working or not.

[Gecko51]

Based on your follow-up, this looks like a known limitation of the current subagent implementation rather than a model string issue.

The model: field in .agent.md does work, but only when the agent is invoked directly. When an agent is delegated to as a subagent from a parent agent session, it inherits the parent session's model instead of using its own model: override. This is consistent with what you're seeing: works fine in a fresh chat, falls back to parent model when called as a subagent.

Whether this is intentional or a bug is a bit unclear. The multi-agent orchestration in VS Code Copilot is still pretty new, and running each subagent on a different model would require separate inference calls with separate billing/quota tracking, which adds complexity. My guess is this isn't fully implemented yet rather than being a deliberate design decision.

A few things worth trying:

• Check if there's a VS Code Copilot extension update that changed this behavior (they've been shipping fast on agent features)
• File a bug report if you want this tracked. The specific reproduction case you have (model works direct, falls back when subagent) is a clean enough repro to be useful
• As a workaround, if you need Claude's behavior for specific steps, start a fresh session with your Claude agent directly rather than delegating from the GPT session