Claude Code's /goals feature automates agent task evaluation

| Feature/Platform | Anthropic Claude Code (/goals)

{ "enrichedKeyPoints": [ "The /goals feature, introduced in Claude Code version 2.1.139 in May 2026, transforms Claude Code from a conversational assistant into an autonomous agent capable of pursuing high-level objectives for extended periods without constant human intervention.", "Unlike traditional single-turn prompts, /goals initiates a persistent agentic loop where a separate, smaller, and faster model (Haiku by default) continuously evaluates a user-defined completion condition after each turn of the execution model, ensuring the task is explicitly satisfied before termination.", "Effective /goal prompts require a clear, measurable end state, a specific verification check (e.g., 'npm test exits 0'), and explicit constraints to guide the agent's autonomous operation.", "This feature is part of Anthropic's broader strategy to enhance 'Hands-Free Productivity' within Claude Code, complementing other tools like the 'Agent View' CLI dashboard for multi-session orchestration and background task execution.", "The use of Haiku as the default evaluator leverages its speed and cost-efficiency, as Haiku 4.5 (released October 2025) offers near-frontier performance on coding and agent tasks at a significantly lower cost and faster speed than larger models." ], "competitorAnalysis": "| Feature/Platform | Anthropic Claude Code (/goals) | OpenAI (General Approach) | LangGraph / Google ADK | MLflow (Evaluation Platform) | Galileo (Agent Reliability Platform) | |---|---|---|---|---|---| | Core Approach | Two-model architecture (execution + dedicated Haiku evaluator) for automated task completion verification. | Model decides when it's done; allows user-defined evaluators to be tagged on. | Independent evaluation possible but requires developers to define critic node, termination logic, and observability. | Complete evaluation platform; LLM judge evaluates full execution traces (tool calls, reasoning, planning). | AI agent reliability platform; 'eval-to-guardrail lifecycle' with Luna-2 (purpose-built small LLMs for evaluation). | | Evaluation Model | Haiku (default), a smaller, faster, cost-efficient model. | Relies on the primary execution model or user-provided evaluators. | Requires developer to define and configure the critic model/logic. | Uses LLM judges and integrates with other eval libraries (Ragas, DeepEval). | Luna-2 (3B and 8B parameter variants), purpose-built small LLMs for evaluation. | | Task Termination | Agent continues until explicit goal condition is met by evaluator. Prevents premature termination. | Model-driven termination; can lead to premature exits if not carefully managed. | Developer-defined termination logic. | Focuses on evaluating the entire execution trace, not just termination. | Aims to prevent failure modes like incorrect tool paths or behavioral drift through continuous evaluation. | | Ease of Use for Evaluation | One-line command for goal definition; default independent evaluator. | Requires custom observability or third-party evaluation logic for robust checks. | Requires significant developer effort to set up evaluation and termination logic. | Comprehensive but requires integration into existing workflows. | Provides pre-built evaluators and auto-tuning, but initial setup for Luna distillation. | | Pricing Model | Subscription plans (Pro, Max) or pay-per-token API. Haiku is cost-efficient. | Typically pay-per-token API. | Varies by underlying LLM and platform. | Varies by platform/tool; often open-source components with enterprise offerings. | Enterprise features require paid tiers. | | Key Use Cases | Autonomous coding tasks, multi-step development, hands-free productivity. | General LLM applications, custom agent development. | Agent orchestration, complex multi-step tasks. | Agent development loop, tracing, prompt optimization, CI/CD integration. | End-to-end agent evaluation, simulation, observability, guardrails for multi-agent systems. |