Claude Code, Anthropic’s coding agent, can now communicate across sessions: one session sends a message directly to another, instead of you copying summaries between terminals by hand. This matters for anyone who develops with Artificial Intelligence (AI), from solo projects to distributed teams. Meanwhile, new tools from Meta, Microsoft, Tencent, and Prime Intellect show where the field is going: agents are being networked, given shared memory, and set up to test their own work.
What is Claude Code?
Claude Code is an AI agent from Anthropic that runs directly in the terminal. Unlike a chatbot, it doesn’t just answer questions — it works in sessions, i.e., separate terminal sessions that previously knew nothing about each other. Inside those sessions, it takes on tasks that used to be manual work: planning changes, writing code, and checking results across large codebases. Claude Code competes with tools such as Codex from OpenAI and Meta’s new Muse Code.
What’s new about team mode?
Since version 2.1.224, Claude Code sessions on macOS and Linux can exchange messages. According to Anthropic, this works in both directions: you can ask another session a question and receive the answer directly. Claude can also send a message on its own, for example when a change affects the work of another session.
The typical use cases are passing on insights, coordinating parallel worktrees, and checking the status of long-running tasks. Worktrees are parallel working copies of the same project; previously, you had to transfer information between them by hand. The message a session sends is a text summary that Claude creates on request.
On the same machine, communication stays local. Between different machines, it runs via Anthropic servers — and there, only answers are possible. The feature is not available on Amazon Bedrock, Google Cloud Agent Platform, or Microsoft Foundry, and administrators can restrict it in the settings.
Pros and Cons of Connected Sessions
Pros:
- Less copy-paste – Insights move directly from one session to another without manual transfer.
- Better parallel work – Sessions coordinate across worktrees and keep track of each other.
- Status checks – One session can ask another about the state of a long task.
- Local option – On the same machine, messages never leave your system.
Cons:
- Limited remote mode – Between different machines, only answers are possible via Anthropic servers, not autonomous messages.
- Missing platform integration – Not available on Amazon Bedrock, Google Cloud Agent Platform, or Microsoft Foundry.
- Control overhead – Because sessions can send messages on their own, administrators need to restrict the feature via settings.
Who benefits from this?
Claude Code is not the only agent that works in teams. As the Swiss trade magazine Netzwoche reports, Meta has introduced Muse Code, a coding agent built on a main agent with specialized sub-agents. The sub-agents stay active for the entire session, keep their knowledge of the repository — the project’s central code storage — and continue certain processes in the background. On larger projects, they work in parallel in isolated Git environments, i.e., separate working copies with version control, without directly touching the developer’s working copy.
The tool also logs all model calls, executed tools, and code changes, which lets it resume a task after an interruption or crash. Meta’s Muse Spark 1.2 model was trained together with Muse Code, including sessions with more than 1,000 tool calls lasting up to 24 hours. According to VentureBeat, this is Meta’s most significant move yet into the coding-agent market. The published benchmarks, however, do not put Muse Spark 1.2 at the top: Claude leads the three evaluations Meta presented. These benchmark figures also come from the provider and are not independently verified.
Teams working in parallel on large codebases benefit most from such architectures. For individual developers, the advantage is smaller, because they keep the context in their heads.
A different approach comes from Prime Agent by Prime Intellect, an open-source harness for coding and research. Here, sub-agents are function calls inside a persistent IPython kernel — a continuously running Python environment — so they run as calls in a live environment rather than as separate processes. The Continual Harness lets the agent edit its own prompts, i.e., working instructions, skills, memory, and sub-agent specifications mid-run. With Opus 5, the company reports 95.5 percent RHAE Best@1 on ARC-AGI-3, just above the stated human expert baseline of 95.4 percent. These figures come from the provider and are not independently verified.
What does shared memory bring?
When several agents work together, the question is where their shared knowledge lives. Tencent Cloud has released TencentDB Agent Memory v2.0 as an open-source memory hub for agent teams. The system turns conversations, documents, and code into four governed, reusable assets: Chat Memory, Skill, LLM-Wiki, and Code-Graph. It is MIT-licensed, self-hosted via Docker (so it can run without a cloud), and integrates with Claude Code, OpenClaw, Hermes, and CodeBuddy.
The differentiator is not retrieval but governance: ACL-based visibility — a list of access rights — decides which agent gets which asset and which version is valid. The MIT license and self-hosting via Docker allow the memory to run in your own infrastructure, without being tied to a cloud. This becomes important as soon as several agents and people work on the same codebase, because it has to be clear which state of knowledge an agent can rely on.
How do you test AI-generated code?
More agents mean more code that needs checking. Microsoft has open-sourced code-testing-generator, a unit-test agent in the MIT-licensed dotnet/skills repository. Before writing anything, it reads the repository: it detects the language, the test framework, existing conventions, and the actual build and test commands. Then it plans, writes, runs, and validates the tests it produces.
On Microsoft’s internal benchmark of 152 tasks, the agent completed 140, compared with 120 for standard GitHub Copilot on the same model. That corresponds to a success rate of 92.1 percent versus 78.9 percent. The gain was concentrated almost entirely in vague prompts and diff-targeted requests. Unit tests check individual functions or modules; a generator that writes them automatically removes exactly the routine work that piles up when AI agents produce a lot of code.
What does this mean for you?
If you’re new to AI-assisted development, the easiest starting point is to work with two Claude Code sessions on the same machine. Open two terminals, have one session send a summary to the other, and check whether the second session picks up the context. The feature works on macOS and Linux, and on the same machine your data doesn’t leave your computer. This gives you a feel for how agents communicate, without having to deal with servers or permissions first.
For more advanced use, combine session communication with a team memory hub such as TencentDB Agent Memory v2.0 so that multiple agents can draw on the same state across projects. Microsoft’s test generator can then verify what the agents produce, and Muse Code’s architecture shows how parallel sub-agents work in isolated environments without disturbing the main working copy. Once more than one agent is involved, keep the access rights in the memory system up to date — that is where team modes succeed or fail.
The new team modes reduce friction: insights travel directly between sessions, sub-agents work in parallel, and shared memory stops every session from starting from zero. But the reported benchmark numbers come from the providers themselves and are not independently verified. The open question is how to organize trust and control in networks of agents that increasingly decide on their own what to share with whom.
Sources
- Claude-Code-Sessions können ab sofort miteinander kommunizieren und Informationen austauschen – The Decoder, 2026-08-08
- Meta lanciert Programmier-KI Muse Code – Netzwoche, 2026-08-07
- Microsoft Open Sources code-testing-generator: A Polyglot Unit-Test Agent That Hits 92.1% Task Completion Versus 78.9% for Stock Copilot – MarkTechPost, 2026-08-06
- Tencent Cloud Open-Sources TencentDB Agent Memory v2.0: A Team-Level Memory Hub for AI Coding Agents – MarkTechPost, 2026-08-07
- Prime Intellect Releases Prime Agent: An Open-Source RLM Harness Where Sub-Agents Are Function Calls Inside Persistent IPython Kernel – MarkTechPost, 2026-08-06
Image: Al Nahian via Pexels