How Claude Code's /goals prevents AI agents from quitting too soon

In AI-driven code migration projects, a green pipeline often masks hidden flaws. Tasks marked complete may still harbor uncompiled code or overlooked dependencies—errors that can take days to surface. This isn’t a flaw in the underlying model; it’s a systemic issue where the agent itself prematurely declares victory. To address this, Anthropic’s latest innovation separates the agent that performs the work from the one that validates its completion.

Claude Code’s /goals feature introduces a dual-model approach: one model executes actions, while another evaluates whether the original goal has been achieved. This separation prevents agents from mistaking partial progress for full completion, a common pitfall in production AI pipelines.

The two-model split explained

AI orchestration platforms from LangChain, Google, and OpenAI have all identified the same challenge: agents terminating tasks before they’re truly finished. Their solutions differ in complexity. OpenAI’s approach allows developers to attach custom evaluators to the default loop, while LangGraph and Google’s Agent Development Kit (ADK) require manual configuration of termination logic and termination nodes. These methods add overhead by introducing new systems to maintain.

Anthropic’s /goals simplifies this by baking an independent evaluator into the workflow by default. Developers define a goal using a clear prompt—for example, "All tests in test/auth pass, and the lint step is clean." Every time the agent attempts to conclude its work, the evaluator model (Haiku by default) checks the task against this condition. If unmet, the agent continues; if satisfied, it logs the achievement and clears the goal. This binary decision—done or not done—enables the lightweight Haiku model to function effectively as the evaluator.

The separation ensures the agent executing the work never conflates progress with completion. Anthropic emphasizes this reduces reliance on third-party observability tools, though enterprises can still integrate them alongside /goals.

Designing measurable completion criteria

For /goals to function effectively, developers must define precise, verifiable conditions. Anthropic recommends structuring goals with three key components:

• A single measurable end state, such as a passing test suite, a clean Git status, or a specific file modification.
• Explicit verification steps, like running npm test with an exit code of 0 or confirming no unintended files were altered.
• Constraints that must remain unchanged, such as prohibiting modifications to unrelated test files.

These criteria ensure the evaluator model has clear, objective benchmarks rather than subjective interpretations of "done."

The broader shift toward auditable AI systems

This innovation reflects a growing trend in agentic AI: the push for stateful, long-running agents that can self-evaluate and improve. Similar evaluator models and verification systems are emerging in reasoning frameworks and coding agents like Devin or SWE-agent. The core principle—separating the builder from the judge—mirrors fundamental software engineering practices, where self-assessment is inherently unreliable.

Sean Brownell, solutions director at Sprinklr, notes that while the two-model loop is sound design, Anthropic isn’t the first to propose it. He highlights that OpenAI and Google introduced comparable features within days of each other, but each took a different philosophical approach to who declares a task complete.

Brownell also cautions that this model excels for deterministic, verifiable tasks—such as migrations, test suite fixes, or backlog clearing—but falls short for nuanced work requiring human judgment, design decisions, or creative problem-solving. In those cases, a human overseer remains irreplaceable.

As AI agents grow more autonomous, the demand for auditable, observable systems intensifies. Features like /goals represent a step toward that future—one where AI doesn’t just perform tasks but proves they’ve been done correctly.