Claude introduces message threads to speed up long AI agent sessions

Solving the history bottleneck

When developers build AI agents, these systems often engage in long-running conversations that can span hundreds of requests. In a standard setup, every time an agent makes a new request, it must resend the entire history of the conversation to the model. This includes every previous message and every result from every tool the agent has used.

As a conversation grows, the amount of data being sent increases significantly. When a session reaches hundreds of thousands of tokens—the basic units of text used to measure data size—the time it takes for the system to process that history adds a delay. This delay happens before the AI even begins to generate its first word of a response.

Anthropic is addressing this issue with a new feature called message threads. According to claude.com, this feature allows agents to send only what is new since the last turn. By removing the need to resend the full history every time, the system can start responding much faster during long sessions.

How the thread system works

The message threads system functions by creating a persistent link between requests. The very first request in a session includes the full conversation history to establish the context. Once that initial request is made, the system generates a unique thread ID that stays with the conversation.

For every subsequent turn, the client only needs to send the new messages along with the ID of the previous response. The system then holds that specific thread in memory. This memory space is encrypted and is not written to a physical disk, which helps maintain the speed and security of the data during the session.

There is a time limit on how long these threads stay active in memory. A thread will expire 5 minutes after the last request is made. Once a thread expires, the client must resend the full conversation history to start a new thread. This ensures that the system does not have to hold onto every single conversation indefinitely.

Performance and token costs

The primary benefit of this new architecture is a reduction in latency, specifically the time to first token (TTFT). For requests involving between 500,000 and 1,000,000 tokens, the median TTFT drops by about 35% when using message threads. This makes the AI feel much more responsive during heavy tasks.

It is important to note that while the amount of data being sent over the wire is smaller, the actual cost of the request does not change. Claude still reads the entire conversation history on every turn to maintain context. Therefore, the total token usage remains exactly the same as it was before the introduction of message threads.

This distinction is important for developers who need to balance speed with their budget. The feature provides a significant performance boost for active sessions where speed is a priority, without changing the underlying pricing structure of the Claude Models API.

Who benefits from this update

This update is specifically designed for high-context workflows where the amount of information carried by an agent is large. One major use case is long-running coding agents. These agents often make hundreds of requests in a single session, and each request might include a new file edit or a new tool result that adds to the total history.

Another primary beneficiary is computer use agents, which are AI systems designed to interact with a screen. With message threads, screenshots can stay with the thread instead of being re-uploaded every time the agent takes a new action. This significantly streamlines the process for the AI as it navigates different interfaces.

The feature also helps with document and image-heavy workflows. Because these types of requests are naturally large, reducing the amount of data sent with each turn makes the overall experience smoother. The more context an agent is required to carry, the more noticeable the reduction in latency becomes.

Availability and next steps

Anthropic has released message threads in public beta across several major platforms. This includes the native Claude Platform as well as the Claude Platform on AWS. It is also available for users on Claude in Microsoft Foundry and Claude Models on Google Cloud, which are managed by Anthropic.

Developers can begin using this feature by adding the thread parameter to their Messages API requests. While the feature is in beta, standard Claude token rates and existing rate limits still apply to all users. This allows for a gradual rollout where developers can test the speed improvements in their own specific environments.

As the technology matures, it will likely become a standard way to interact with large language models. For now, it provides a way for developers to build more complex, multi-step agents that can handle massive amounts of data without the performance degradation that usually comes with long conversations.