What happens when an agent can ask another agent for help?
That’s exactly the idea behind one of the most interesting features of Qwen Code: support for ACP (Agent Communication Protocol), a protocol designed so that different agents can collaborate with each other.
Instead of competing.
They cooperate.
And that small paradigm shift could end up being much more important than the next model with a few extra points on a benchmark.
What is ACP?
ACP stands for Agent Communication Protocol.
Its goal is quite simple:
Allow an agent to delegate a task to another specialized agent.
Instead of trying to solve everything by itself.
Think of it as a microservices system.
Each service does one thing very well.
Agents start to function in a similar way.
From universal agent to specialized agent
Until now the typical workflow was this:
Claude Code
↓
Does everything
Plans.
Writes code.
Reviews.
Documents.
Executes commands.
Everything goes through the same agent.
ACP proposes a different approach.
Claude Code
↓
Qwen Code
↓
Specialized result
Each agent can take on a different role.
A practical example
Imagine you’re working on a huge project.
You ask Claude Code:
“Refactor the authentication module and verify that it doesn’t break the permissions system.”
Claude could decide:
-
to plan the work,
-
to understand the context,
-
to identify the files involved.
Then delegate:
“Qwen, analyze only the dependencies and possible regressions.”
Qwen performs that specific task.
Returns the result.
Claude continues with the implementation.
For the developer, the experience can feel like there’s a single agent.
But behind the scenes, several are collaborating.
Why do this?
Because no agent is best at everything.
One can excel at:
-
planning,
-
long context,
-
architecture.
Another can be excellent at:
-
code generation,
-
mathematical reasoning,
-
speed,
-
cost.
ACP allows you to leverage the strengths of each.
A hybrid stack
This fits perfectly with a trend we’ve been seeing over the last few months.
Many developers already work with:
-
Claude Code
-
Aider
-
Codex
-
OpenCode
Not because one is better.
But because each brings something different.
ACP simply takes that idea one step further.
Now the agents themselves can coordinate.
Installation
Qwen Code can be installed via npm:
npm install -g @qwen-code/qwen-code
Then start it:
qwen
To use ACP you need a recent version of Qwen Code and an agent compatible with the protocol.
Configuration depends on the environment (Claude Code, Codex, or other compatible clients), so it’s best to follow the instructions in the official repository as support continues to evolve.
What tasks make sense to delegate?
Some interesting examples:
Dependency analysis
A specialized agent can review relationships between modules while the main one continues working.
Code review
After generating an implementation:
Claude
↓
Qwen reviews
↓
Claude fixes
Documentation
While one agent modifies the code.
Another can generate:
-
documentation,
-
changelog,
-
comments,
-
examples.
Testing
One agent can generate the tests.
Another execute fixes.
Another review coverage.
Each focused on a specific task.
Does this replace Claude Code?
No.
And that’s exactly the point.
ACP doesn’t try to replace any agent.
It tries to connect them.
It’s an important distinction.
For a long time we thought the future was about finding the perfect agent.
There are increasingly more signals that one probably doesn’t exist.
What does seem to be emerging is an ecosystem of specialized agents that collaborate with each other.
What’s really interesting
A few months ago we compared models.
Today we’re starting to compare ecosystems.
Because the competitive advantage no longer depends solely on the model you use.
It also depends on:
-
what tools it can invoke,
-
what memory it shares,
-
what MCP servers it uses,
-
what other agents it has available.
Agents are starting to look less like isolated applications.
And more like work teams.
It wouldn’t be surprising if within a year working with a single agent became as unusual as developing a modern application using a single service.
Maybe the future isn’t an extremely intelligent agent.
Maybe it’s a group of reasonably good agents…
…that know exactly when to ask for help.
Repository: GitHub - QwenLM/qwen-code: An open-source AI coding agent that lives in your terminal. · GitHub