Coding agents increasingly work as connected groups that exchange findings, share knowledge, and divide tasks among themselves. This affects people who maintain larger software projects as well as those using Artificial Intelligence (AI) to create small internal applications. New features from Anthropic, Meta, Tencent, and Cloudflare show how the individual chat window is becoming a coordinated workspace.
What defines a coding agent team?
A coding agent is an AI system that can go beyond answering programming questions to plan work, modify code, call tools, and check results. Until now, multiple agents often operated separately: one investigated a defect and another wrote tests, while a person copied information between their sessions. Anthropic is trying to close that gap with a new Claude Code feature.
Starting with version 2.1.224, Claude Code sessions on macOS and Linux can communicate with each other. One session can send another a text summary, ask a question, or independently report that a change affects the other session’s work. According to Anthropic, possible uses include checking the status of long-running tasks and coordinating parallel worktrees, which are separate working areas of the same Git project.
Communication runs locally when the sessions are on the same computer. Between different machines, it goes through Anthropic’s servers and is limited to replies, according to the report. Administrators can restrict the feature through settings, and it is unavailable on Amazon Bedrock, Google Cloud Agent Platform, and Microsoft Foundry.
Meta uses a similar general idea in Muse Code but organizes the collaboration within one system. A main agent delegates work to specialized subagents that remain active throughout the session. Meta says this allows them to retain knowledge about the repository, the central storage location for a software project, and continue certain operations in the background.
Why shared memory matters
Communication alone does not turn several agents into a dependable team. They also need to know which information is valid, who is allowed to use it, and which version of a file or rule is current. Without that control, adding agents can multiply confusion just as easily as labor.
Tencent Cloud has introduced TencentDB Agent Memory 2.0 as an open-source memory hub. It is intended to convert conversations, documents, and code into four reusable assets: Chat Memory for conversational knowledge, Skill for learned procedures, LLM-Wiki for organized knowledge collections, and Code-Graph for relationships within code. LLM stands for large language model.
According to the provider, the main distinction is not simply retrieving information but governing it. Access control lists determine which agent can see a particular asset and which version is valid. The software uses the MIT license, can be self-hosted through Docker, and is designed to integrate with Claude Code, OpenClaw, Hermes, and CodeBuddy.
This approach is particularly relevant when several groups in an organization use the same agents. A support agent could reuse a documented internal procedure, for example, without automatically gaining access to every project document. A coding agent could refer to an approved architecture decision instead of relying on an outdated answer from an earlier chat.
The sources do not identify separate availability or specific language support for Switzerland. Still, the distinction between local communication, provider-operated servers, and self-hosted memory matters to Swiss organizations deciding where source code and internal documents should be processed. Whether a particular setup meets applicable privacy requirements depends on the individual use case; the reports do not provide a definitive assessment.
What makes larger tasks possible?
Meta presents Muse Code as an agent for extensive changes across large codebases. The system is supposed to plan modifications, write code, and verify the result. Specialized subagents can work in parallel inside isolated Git environments without changing the developer’s primary working copy directly.
This could help when a team needs to investigate a faulty feature, add suitable tests, and update documentation at the same time. Instead of coordinating three separate chats, a main agent can distribute the subtasks and combine their findings. Whether Muse Code can perform these workflows reliably in practice has not been independently verified.
Meta also says the tool records every model call, tool execution, and code change. These logs are intended to let it resume a task after a crash or interruption. The underlying Muse Spark 1.2 model was reportedly used in sessions involving more than 1,000 tool calls and lasting up to 24 hours, but these figures come from the provider.
Agent teams are not limited to source code. Cloudflare’s Kitesurf is a cloud-hosted browser for AI agents that is designed to navigate websites, fill out forms, and perform other browser-based tasks. Unlike a human browser, it has little need for themes or visible tabs, but it must manage context windows, computing performance, token costs, and scale; tokens are small units of text processed by a language model.
Kitesurf is free during its beta period and runs on Cloudflare Workers. Cloudflare claims it uses less processor power and memory than Chromium for common agent tasks such as taking screenshots and extracting Hypertext Markup Language (HTML). This efficiency claim comes from the provider and has not been independently confirmed in the supplied sources.
One practical example would be an agent retrieving information from a web application and then completing a form while a second agent adjusts the related program logic. That broader reach also increases the attack surface. Cloudflare specifically mentions prompt injection, an attack in which manipulated content on a website gives the agent harmful instructions.
Pros and Cons of connected coding agents
Pros:
- Parallel work – Multiple agents can handle defect analysis, code changes, and tests at the same time rather than completing each task in sequence.
- Less manual copying – Sessions can exchange summaries, questions, and status updates directly.
- Persistent context – Long-lived subagents and shared memory can reduce the loss of project knowledge between individual work steps.
- Traceability – Logs covering model calls, tools, and changes can show how a result was produced and where a task stopped.
Cons:
- Cascading errors – One incorrect assumption can pass from one agent to others, making it harder to identify.
- Broader permissions – Agents working with code, documents, and websites need narrowly defined access and isolated work areas.
- Provider dependence – Features vary by operating system, platform, and cloud service; Claude Code session communication, for example, is missing from several managed platforms.
- High resource use – Long tasks and multiple subagents generate far more model calls than a simple chat response.
Cloudflare OS illustrates why these limits matter. It is an open-source platform for so-called vibe coding, where people describe an application in natural language instead of writing all of its code themselves. Cloudflare reports that thousands of its employees use the system daily to create documents and presentations, automate repetitive work, and build small applications that visualize data.
The platform separates individual applications and even individual documents into their own sandboxes, meaning isolated execution environments. Agents begin without resource permissions and must request access, while outbound networking for server code is disabled by default. Cloudflare’s broader assertion that the setup prevents agents from introducing significant security flaws remains a provider claim rather than a guarantee.
What this means for your work
If you are getting started, one agent with a clearly limited assignment remains the most sensible first step. You could ask it to investigate a contained defect and propose suitable tests rather than immediately letting it process an entire project autonomously. Review its code changes, tool calls, and file access before adding more agents.
Advanced users can gain more by assigning distinct roles. One agent might plan the task, a second modify a specific part of the program, and a third review tests or documentation. Shared memory should be versioned and divided by permission, while isolated working copies are easier to control than direct changes to production systems.
Small applications made by nonprogrammers can benefit from the same separation. An employee could describe a work-data visualization app in natural language while the agent receives access only to explicitly approved datasets. Cloudflare OS provides one possible framework, but it does not prove that every application created this way will automatically be secure or factually correct.
Energy use also belongs in the decision. An analysis of climate scientist Zeke Hausfather’s Claude Code activity tracked about 1,100 inputs, more than 14,000 model calls, and 3.2 billion tokens over eight weeks. Estimated electricity use reached roughly 170 kilowatt-hours, or around 150 watt-hours per input.
That intensive agent activity was about 600 times higher than the published figures for a simple text prompt: Google reports 0.24 watt-hours for a median Gemini prompt, while OpenAI gives 0.34 watt-hours for an average ChatGPT request. The comparison is only an approximation because people outside AI labs do not have reliable figures for the energy consumed per token by a leading model. Projected across a year, the estimated emissions in this individual case were similar to those of an electric clothes dryer.
Agent teams can distribute larger software tasks more clearly and preserve knowledge better than isolated chat windows. Their value depends less on the number of agents than on defined roles, access rules, logs, and isolated workspaces. The open risks are the reliability of long autonomous workflows, security when agents access external content, and energy demand when a handful of inputs turns into thousands of model calls.
Sources
- Claude-Code-Sessions können ab sofort miteinander kommunizieren und Informationen austauschen – THE DECODER, 2026-08-08
- Tencent Cloud Open-Sources TencentDB Agent Memory v2.0: A Team-Level Memory Hub for AI Coding Agents – MarkTechPost, 2026-08-07
- Meta lanciert Programmier-KI Muse Code – Netzwoche, 2026-08-07
- Cloudflare launches Kitesurf, a browser built for AI agents – TechCrunch, 2026-08-07
- Cloudflare open-sources vibe-coding platform for people who aren’t coders – Ars Technica, 2026-08-06
- KI-Agenten verbrauchen etwa 600-mal mehr Energie als ein einfacher Chat-Prompt – THE DECODER, 2026-08-08
Image: Bhavishya 🙂 via Pexels