Coding agents are starting to handle work as a coordinated team rather than as a single chat. This affects you if you use Artificial Intelligence (AI) to build small applications, modify existing software, or maintain a large codebase. New tools from Anthropic, Meta, Microsoft, Tencent, and Cloudflare illustrate how agents may divide work, knowledge, and testing among themselves.
What is changing about coding agents?
Starting with version 2.1.224, Anthropic allows separate Claude Code sessions on macOS and Linux to communicate. According to the report on cross-session communication, one session can send another a text summary, ask a question, or receive an answer directly. Previously, you often had to copy this information manually between terminal windows.
A session can also send a message on its own when a change affects another session’s work. Anthropic lists sharing findings, checking the status of long-running tasks, and coordinating parallel worktrees as typical uses. A worktree is a separate working copy of the same software project, allowing changes to proceed in parallel without immediately overwriting one another.
This is not yet a fully self-managing digital development team. It does, however, create a connection between agent runs that were previously isolated. Instead of giving one agent an extremely long list, you can assign distinct roles to several sessions and let them exchange results.
Communication on the same computer runs locally. Communication between different computers uses Anthropic’s servers, according to the report, and only replies are supported in that case. Administrators can restrict the feature through settings, and it is unavailable on Amazon Bedrock, Google Cloud Agent Platform, and Microsoft Foundry.
How do agents divide work and knowledge?
Meta is pursuing a similar but more integrated approach with Muse Code. According to the company, the Muse Code programming agent combines a primary agent with several specialized subagents. They remain active throughout a session, retain knowledge about the repository—the central storage location for a software project—and decide when to report results to the primary agent.
In a larger project, one subagent could examine the existing codebase while another prepares a change and a third checks the results. Meta also describes parallel work inside isolated Git environments. Git is a widely used system for tracking file changes and managing different versions of a project.
Muse Code records model calls, tool use, and code changes, according to Meta. This is intended to let it resume a task after an interruption or crash. Meta also reports training sessions involving more than 1,000 tool calls and lasting up to 24 hours; these are provider claims and were not independently verified in the supplied sources.
Dividing tasks is not enough if each agent must repeatedly rediscover the same context. TencentDB Agent Memory 2.0 is designed to turn conversations, documents, and code into reusable knowledge assets. The open-source memory hub for agent teams distinguishes between chat memory, reusable skills, a language-model wiki, and a code graph that maps relationships within a codebase.
According to the summary, its focus is not simply retrieving information but governing it. Access rules determine which agent can see each asset and which version is valid. The system uses the MIT license, can be self-hosted through Docker, and integrates with tools including Claude Code; the source provides no evidence about its reliability in everyday use.
How are the results checked?
As several agents produce code more quickly, the volume of work requiring review also increases. Microsoft has released an agent for generating unit tests. Unit tests are automated checks of individual software components, intended to determine whether they produce the expected result under defined conditions.
Microsoft’s code-testing generator first reads the repository and detects the programming language, testing framework, existing conventions, and the commands actually used to build and test the software. It then plans, writes, runs, and validates the tests. This gives it a defined role alongside agents responsible for analysis and implementation.
In Microsoft’s internal comparison, the tool completed 140 of 152 tasks, while standard GitHub Copilot completed 120 using the same model. Those results correspond to the 92.1 percent and 78.9 percent completion rates stated in the source title. The summary says the advantage was concentrated in vague prompts and requests targeting specific changes; because this was an internal benchmark, it cannot automatically be generalized to your project.
A passing test is not proof that software has no defects. Agents may write tests that miss important edge cases or merely confirm the behavior already present. The useful division of labor is therefore not “one agent writes, another checks, job done,” but “agents produce reviewable intermediate results whose meaning a person evaluates.”
Pros and Cons of agent teams
Pros:
- Parallel work – Separate sessions can analyze, implement, and test at the same time when their responsibilities are clearly divided.
- Less manual copying – Claude Code sessions can exchange summaries and answers directly instead of making you relay everything between terminal windows.
- More context – Persistent subagents and shared knowledge assets may reduce the need to rediscover the same project details.
- Specialized checks – A dedicated testing agent can read existing conventions and execute its tests rather than merely suggesting test code.
Cons:
- New error chains – One incorrect assumption can be passed to several agents and acquire the appearance of confirmation.
- Harder oversight – Parallel sessions, background tasks, and long tool chains make decisions more difficult to trace.
- Data risks – Server-based communication and shared memory require clear rules for source code, documents, and access permissions.
- Unclear total cost – The sources provide no usage prices, and open-source components do not eliminate model, hosting, or human review costs.
Cloudflare shows that this development is not limited to professional software developers. The company has open-sourced a platform that turns natural-language descriptions into applications, which it tested internally with people who are not software engineers. Cloudflare says employees use it to create documents and presentations, automate repeatable work, and build small applications for visualizing data.
Cloudflare aims to limit risk through separate sandboxes. A sandbox is an isolated runtime environment that restricts a program’s access to other data and systems. Agents begin without permissions and must request access, while outbound networking for server code is disabled by default. The broader claim that this makes serious security bugs practically impossible comes from the company itself.
What does this mean for your work?
If you are getting started, you should not immediately release five agents onto an existing business application. A sensible first step is a small, separate project with two roles: one session creates a narrowly defined change, while another examines the code or runs tests. Write down the assigned task, permitted files, and expected result so you can recognize deviations.
Simon Willison’s Raccoon Heist game experiment offers a concrete example. He supplied Claude Fable 5 with older screenshots and a description, then had it create a playable project in Claude Code for web. Because testing ongoing work there was difficult, he used GitHub Pages and asked the agent to commit an initial web page early; the process shows why quick, visible checkpoints remain necessary even when a one-shot result looks impressive.
If you are more experienced, you can define roles more tightly: one agent maps the codebase, one works in an isolated Git environment, one generates tests, and another summarizes unresolved risks. More agents do not automatically produce more value; clear handoffs, limited permissions, and verifiable results matter more. Shared memory becomes particularly useful when several sessions need the same project rules, documents, and earlier decisions.
The sources provide no Switzerland-specific information about regional availability, supported languages, or data locations. Swiss companies, public agencies, and educational institutions therefore still need to determine whether a particular service processes source code and documents locally or sends them through foreign servers. Local communication between Claude Code sessions and self-hosted components such as TencentDB Agent Memory may be relevant, but neither replaces checking the actual configuration and access paths.
Coding agents working as a team are more than a larger chat window: they combine specialization, parallel work, shared knowledge, and automated testing. The practical gain is most likely to appear in clearly separable tasks and large codebases, while small projects may accumulate more coordination than benefit. The unresolved risk is whether agents can detect one another’s errors and false assumptions rather than merely passing them along more efficiently.
Sources
- Claude-Code-Sessions können ab sofort miteinander kommunizieren und Informationen austauschen – Unknown, 2026-08-08
- Tencent Cloud Open-Sources TencentDB Agent Memory v2.0: A Team-Level Memory Hub for AI Coding Agents – Unknown, 2026-08-07
- Meta lanciert Programmier-KI Muse Code – Unknown, 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 – Unknown, 2026-08-07
- Cloudflare open-sources vibe-coding platform for people who aren’t coders – Unknown, 2026-08-06
- One-shotting a Raccoon Heist game using Claude Fable 5 – Unknown, 2026-08-05
