Major Claude Code Update: A Complete Guide to Agent View and the Goal System

Overview

Anthropic recently released a series of major updates for Claude Code, with the most notable being the brand-new Agent View and Goal system. These features mark Claude Code's evolution from a terminal coding assistant toward a full-fledged agentic development management platform. This article provides an in-depth analysis of the core value and usage of these new features.

Agent View: Centralized Agent Management

From Terminal Assistant to Orchestration Platform

Agent View is the most significant workflow change in this update. It transforms Claude Code into a centralized agent management system — you get a unified dashboard where you can monitor all active Claude Code sessions in real time.

Agent Orchestration refers to the technical architecture for coordinating multiple AI agents working together. In traditional development tools, each AI assistant session operates independently, requiring developers to manually switch between multiple terminal windows. The core value of an orchestration platform lies in providing a unified control layer — similar to how Kubernetes manages containers — enabling multiple autonomously running agents to be centrally monitored, scheduled, and intervened upon. The introduction of Agent View marks Claude Code's evolution from a single-conversation tool to a development platform with "command center" capabilities.

With Agent View, developers can run multiple coding tasks simultaneously without each one occupying a separate terminal tab. Each Agent runs independently in the background, while you manage everything from a single interface.

Specifically, you can intuitively see:

• Which agents are currently running
• Which are waiting for human feedback
• Which tasks have been completed

How It Works

Agent View's operational logic is quite intuitive:

1. Overview all processes: Press the left arrow key in any active session, or run the claude agent command directly in the terminal to view all running Agents
2. Quick preview and reply: Without leaving your current workflow, select a session preview to directly respond to agents awaiting decisions
3. Seamless switching: Press Enter to access the full conversation history and quickly jump between sessions

Background Session Functionality

Claude Code now supports fully background sessions. For any existing workflow, you can transfer it to Agent View via the BG command, or use the claude bg command directly to skip the foreground interface and input a task, immediately launching a new autonomous session for parallel processing.

However, it's important to note that running multiple agents simultaneously consumes a significant amount of tokens. Tokens are the basic unit of measurement for how large language models process text — typically about 4 English characters or 1.5 Chinese characters correspond to 1 token. When multiple agents run in parallel, each Agent independently consumes input/output tokens, with costs scaling linearly. Taking Claude 3.5 Sonnet as an example, API pricing is approximately $3/million tokens for input and $15/million tokens for output. A complex background coding task might consume tens of thousands to hundreds of thousands of tokens, and with multiple Agents running simultaneously, hourly API costs could reach several dollars or more. This feature is therefore particularly suited for:

• Long-running autonomous agents (e.g., PR monitoring)
• Updating dashboards
• Handling recurring background tasks

This feature is currently in research preview and is available to Pro, Max, Team, Enterprise, and Claude API users.

Goal System: Result-Oriented Autonomous Execution

Moving Beyond Step-by-Step Confirmation

The Goal system is another major feature in this update, with its design philosophy drawing inspiration from OpenAI Codex's goal-directed execution system. It introduces a persistent result-oriented execution mechanism natively built into Claude Code.

OpenAI Codex (especially its CLI version) introduced the concept of "goal-driven execution": developers describe the desired end state, and the model autonomously plans and executes multi-step operations until the goal is achieved, rather than waiting for instructions at each step. This stands in stark contrast to the traditional REPL (Read-Eval-Print Loop) interaction model, which requires users to actively intervene at every step. Claude Code's Goal system further distinguishes three execution paradigms on this foundation, reflecting the paradigm shift from "conversational AI" to "autonomously executing AI."

The core change is this: instead of manually prompting the model at every step, you can now set completion conditions directly, and Claude will autonomously work through multiple rounds until the goal is achieved.

Use Cases and Workflow Mode Comparison

The Goal system works in any scenario — whether in interactive sessions, CLI mode, or remote control. The agent no longer needs to constantly wait for user commands or stop for confirmation at every step.

Anthropic has made clear distinctions between different working modes:

Mode	Behavior
Goal	Executes continuously until a specific result is achieved
Loop	Runs based on time intervals, stop hooks, or waits for external conditions
Regular Chat	Waits for manual user input after each response

These three modes cover the complete automation spectrum from fully autonomous to fully interactive, allowing developers to flexibly choose based on the risk level and complexity of the task.

Mobile Remote Control Support

A noteworthy highlight is that Anthropic now allows mobile access to Claude Code. This means you can start long-running autonomous tasks on your phone, set a goal for Claude, let it run overnight, and check the results when you return.

Prompt Compression Optimization: Enhanced Long-Session Experience

Beyond the two major features above, Claude Code has also updated its system prompt compression approach.

The necessity of Prompt Compression stems from the context window limitations of large language models. For example, Claude 3.5's context window is 200K tokens, and during extended coding sessions, conversation history continuously accumulates until it eventually exceeds the window limit. Earlier compression implementations often led to "context drift" — where the model forgets important early instructions or constraints (such as code style requirements or architectural constraints), resulting in inconsistent behavior. The core of this optimization establishes an information priority mechanism that reduces historical context through summarization and redundancy removal while ensuring the user's core intent is preserved with priority.

Key improvements include:

• Silent compression: No more obvious pop-up notifications; the process is now seamless
• Priority preservation of key instructions: The compression prompt itself has been optimized to ensure the model prioritizes retaining important user intent during compression
• Reduced context drift: Helps maintain consistency during long-running agent workflows

This is a substantial experience improvement for developers who frequently engage in extended coding sessions, especially when running autonomous tasks for hours in Goal mode, where maintaining context consistency is particularly critical.

Summary and Outlook

Anthropic's series of updates represents a qualitative leap in Claude Code's agent orchestration capabilities. Agent View solves the pain points of multi-task management, the Goal system makes autonomous execution possible, and prompt compression optimization ensures quality in long sessions.

However, token consumption and usage quota limitations remain the primary bottlenecks. While multi-agent parallel execution is powerful, it also multiplies quota consumption. We look forward to Anthropic further raising usage quota limits in future versions, allowing these powerful features to truly reach their full potential.

Key Takeaways

• Agent View upgrades Claude Code from a terminal coding assistant to a centralized agent orchestration management platform, supporting multi-task parallelism and unified dashboard management
• The Goal system draws from Codex's goal-driven execution philosophy, enabling result-oriented autonomous multi-round execution without step-by-step manual confirmation
• New mobile support allows setting long-running autonomous tasks on your phone and remotely monitoring execution results
• The prompt compression system has been optimized to silent mode, preserving key instructions and user intent through an information priority mechanism, reducing context drift in long sessions
• Token costs scale linearly when running multiple agents in parallel, and usage quota limitations remain the primary bottleneck