Artificial intelligence company Anthropic announced on Friday (August 8) that its AI coding tool, Claude Code, has received a major feature update: AI agents across different sessions can now send messages to each other. This means that AI agents running in multiple terminal windows opened by a developer can proactively align on progress, share key discoveries, and even automatically warn about code conflicts—completely eliminating the era of manually copying and pasting context between different windows.

Simultaneously, Anthropic also announced that starting August 14, for Pro, Max, and Team subscribers, Claude Code’s default permission mode will shift to Auto Mode, which uses an independent AI classifier to intercept risky operations in real time. Together, these two updates mark Anthropic’s full-scale evolution in the AI development ecosystem from “single-agent assistance” to “secure multi-agent collaboration.”

Cross-Session Messaging: Letting AI Collaborate Like Team Members

According to Anthropic’s official documentation, the cross-session messaging feature requires Claude Code v2.1.224 or higher and currently runs on macOS and Linux systems. When a session meets the version requirements, the feature is enabled by default with no additional setup needed.

The launch of this feature directly addresses the biggest pain point in multi-task parallel development: context synchronization. Previously, if a developer had multiple terminal windows open simultaneously—handling frontend, backend, or different Git Worktree branches—when one session reached a key conclusion or completed a dependency, the developer had to manually copy and paste context into another window. This was not only inefficient but also prone to missing critical information.

Now, a user simply needs to tell Claude something like “Notify test-session about this API interface modification,” and Claude will automatically compile a summary and send it. Even smarter, when Claude detects that a modification it just made to the codebase will affect another running session, it will proactively send an alert to that session without any user intervention.

From a technical implementation standpoint, Claude uses two tools to accomplish this functionality: ListAgents, used to find reachable agents, and SendMessage, which sends messages to a specified agent by name. The receiving Claude only reads messages during the gaps between tool calls within a work turn, so running tools are not interrupted. When the receiving session is idle, Claude Code uses the message to initiate a new turn.

Anthropic outlined four major application scenarios in its official documentation:

Handing off investigation results: When one session discovers a significant change or makes a decision, Claude summarizes it for sessions handling affected areas, rather than requiring the developer to re-explain.
Coordinating parallel work trees: When sessions are working on the same repository in different workflow trees, Claude can tell other sessions what has been completed.
Getting status on long-running work: Allowing a migration or test run to report progress to the session the developer is watching, or proactively inquiring from that session.
Cross-machine replies: Replying to messages received from sessions on another machine or network. However, for cross-machine replies, Claude can only reply to messages and cannot proactively initiate message exchanges.

It is worth noting that cross-session messaging is not all-powerful. According to tech media outlet 9to5Mac, the feature does not apply to approval permission requests or configuration changes. Commands such as /compact will arrive as plain text rather than actual commands; if acting on a message requires permissions, the receiving session will still prompt the user for approval.

Users can see which sessions Claude can access by running the /list-agents command. The list includes: child agents running within the current session, other local sessions on the same machine (including background sessions), and, when connected via remote control, Claude Code sessions on other machines or on the web. However, Claude cannot send messages to remote sessions to proactively initiate a conversation; it can only reply.

Auto Mode: Using AI to Guard AI’s Safety Boundaries

Alongside the launch of cross-session communication, Anthropic also announced another important security update.

According to IT Home, Anthropic issued a notice stating that starting August 14, for Pro, Max, and Team subscribers, Claude Code’s default permission mode will be adjusted from traditional manual approval to Auto Mode.

The core of Auto Mode is an independent AI classifier that evaluates every tool call and Shell command in real time: allowing safe operations while automatically blocking destructive, irreversible, or unauthorized actions.

Anthropic released a set of comparative test data highlighting Auto Mode’s effectiveness. In tests conducted with 1,053 paying beta testers, the traditional manual approval mode identified only 13.6% of dangerous commands, while Auto Mode was able to identify 89% of dangerous commands.

In defense tests against prompt injection attacks, the results were even more striking. Security assessment firm Trajectory Labs used 72 scenarios, running 10 tests per scenario for a total of 720 attack attempts, covering Claude Code v2.1.205 and Codex v0.144.5. The test results showed: Claude Fable 5, Opus 5, and Sonnet 5 all blocked attacks completely in Auto Mode, achieving a 0% attack success rate. GPT-5.6 Sol had a 5.83% success rate under Codex’s Auto-review mode and a rate as high as 19.03% under Full Access mode.

Anthropic also noted that the bypassPermissions mode without additional safeguards had an average attack success rate of 0.09%, but the new Auto-review version OpenAI released last week could change this result.

In terms of coverage, Claude Enterprise, Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud’s Agent Platform, and Microsoft Foundry currently remain in an opt-in state. Anthropic plans to roll out adjustments across these platforms within the next month, setting Auto Mode as the default option and removing associated overhead.

A New Paradigm for Multi-Agent Collaboration

Viewed together, these two updates show Anthropic building a more complete collaboration framework for AI coding tools. Cross-session messaging breaks the current limitation of AI Code Agents working in isolation, advancing Claude Code into a new paradigm of multi-agent coordination. It not only eliminates the friction of context transfer but also saves token consumption in long-task scenarios, enhancing collaboration efficiency between AI agents.

Auto Mode, meanwhile, provides a safety guarantee for this multi-agent collaboration. When multiple AI sessions can autonomously communicate and send instructions to each other, an automated security layer capable of identifying and intercepting dangerous operations in real time becomes indispensable.

This series of moves by Anthropic also reflects how competition in the AI coding tool arena is shifting from pure code generation capability toward more complex engineering collaboration capabilities and security governance capabilities. When developers can manage multiple AI agents like managing a real development team—assigning tasks, aligning progress, and issuing mutual warnings—the role of AI in software development is evolving from tool to collaborator.

As Anthropic stated in its official documentation: “When one of your sessions gets information that another session needs, Claude passes it along automatically—no copying and pasting between terminals required.” This perhaps signals that the efficiency of AI development teams is about to experience a qualitative leap.