Anthropic’s Opus 5.5: Why Your AI Agent Stops Mid-Task

Developers deploying AI agents, particularly with Anthropic’s new Opus 5.5 model, have encountered a puzzling issue: their AI assistants stop working mid-task, only to report progress and then go silent. This behavior, often described as an AI “clocking out” prematurely, has led to frustration, but Anthropic reveals the cause isn’t AI laziness, but rather how older agent programs interpret the model’s overly communicative nature.

The “Too Eager to Report” Problem

The core of the issue lies in Opus 5.5’s enhanced ability to provide progress updates and summaries. While this was a key selling point – promising more proactive communication – it clashes with the rigid logic of many existing agent frameworks. These frameworks often interpret an AI’s signal of having finished its current turn (an end_turn API response) as the completion of the entire task, especially if the AI isn’t actively calling a tool in that moment.

Anthropic acknowledges this “halfway dashing off” behavior in its updated prompt guide. The model isn’t intentionally stopping; it’s simply reporting its progress, and older systems mistakenly treat this report as the final deliverable. The official guidance states, “A plain text end of turn should be viewed as a report, not as proof of task completion.” This means the AI might be ready to continue, but the application has already declared the job done.

Four “Good Student” Habits Causing Trouble

Anthropic identifies four specific scenarios stemming from Opus 5.5’s advanced capabilities that can lead to premature task halts:

• Abstract Planning: The AI might generate a lengthy summary of next steps but fail to execute any actual tools or actions, leaving the future tasks perpetually on paper.
• Excessive Politeness: The model may pause its work to ask for confirmation on the next step, “If you don’t mind, I’ll proceed with X,” effectively halting until a human responds – a problem for unattended agents.
• Simulated Request for Approval: The AI could list decisions needing human sign-off, even if these decisions don’t actually block its progress, creating unnecessary bottlenecks.
• Report Compulsion: The model might halt simply because it feels it has reached a natural stopping point for a progress report or has completed a small sub-task, regardless of the overall task status.

Ironically, these behaviors are byproducts of features designed to make the AI more helpful and transparent. When integrated into older, less flexible agent architectures, these “good habits” inadvertently trigger task termination.

Anthropic’s Three Solutions for Uninterrupted AI

To address these issues and ensure AI agents complete their tasks, Anthropic proposes three main strategies:

1. Task Checklists: Break down large tasks into smaller, actionable items managed by a to-do list tool or text. The application should automatically prompt the AI to continue if the checklist isn’t cleared and the AI hasn’t explained any blockers. An example prompt is: “Your task list still has unfinished items: migrate the remaining two endpoints and update their tests. Continue working. If you are blocked on any item, state where the blockage is.”
2. Strict Acceptance Testing: Define clear completion criteria beforehand. A separate, potentially smaller model, can then review the AI’s output each turn against these standards. If the criteria aren’t met, the reasons for failure are fed back to the AI for rework.
3. Hard Stops: Implement a system that forces a human review if an automated task fails to progress after a set number of retries (e.g., two or three). This prevents wasted API credits on stuck tasks.

Additionally, prompt engineering is crucial. Anthropic provides system prompts that explicitly forbid the problematic stopping behaviors while clearly defining when an AI is genuinely allowed to pause (e.g., when blocked by user input or protected resources).

Navigating API Changes and Hidden Pitfalls

Migrating from Opus 5 to Opus 5.5 also involves understanding several API changes that can cause 400 errors if not handled correctly:

• The thinking parameter can no longer be disabled and must be set to adaptive, with the effort parameter controlling cognitive depth.
• Forcing tool calls via tool_choice is deprecated; auto is recommended, with tools specified in the prompt.
• thinking blocks are now tied to model and context; modifying system prompts or messages within a thinking block may cause errors for accounts created after August 31, 2026.
• Older versions of the computer operation toolset are deprecated on Claude API and Google Cloud, requiring an update to computer_toolset_20260801.

Beyond explicit errors, several hidden pitfalls can affect agent performance:

• Invisible Work: Text between tool calls, previously visible as regular text, is now part of the thinking block, which is omitted by default. This can make long tasks appear silent and stuck, even if the AI is working. Setting display to updates or summarized can reveal progress.
• Answer Truncation: The max_tokens limit now applies to both thinking and text, potentially cutting off responses prematurely if not adjusted.
• Token Costs for Hidden Thoughts: Even if not returned to the user, thinking processes consume tokens and incur costs.
• Result Parsing: Developers must distinguish between thinking and text blocks in responses and ensure that the AI’s thinking history is passed back accurately between tool calls.

Tuning Costs and Performance with Effort Levels

With thinking being mandatory, the effort parameter (ranging from low to max) becomes the primary control for cost and performance. Anthropic notes that Opus 5.5’s medium setting can match or exceed Opus 5’s high setting, offering significant cost savings for simpler tasks on lower settings. However, at higher settings, Opus 5.5 performs much more extensive reasoning per turn than its predecessor, potentially increasing token usage and turn length.

A practical recommendation is to start with the medium effort level, test with actual data, and only increase it if necessary. Lowering the effort level is more effective than relying solely on prompt instructions to reduce unnecessary thinking.

Eliminating “AI Flavor” with Blacklists

For frontend development, relying on generic instructions like “avoid AI-like design” often results in predictable, template-based interfaces. Anthropic suggests a more effective approach: using explicit blacklists. By specifying undesirable elements—such as cream or off-white backgrounds, italicized emphasis in titles, 01/02 numbering, monospace tags, or capsule buttons—developers can guide the AI away from generic “AI aesthetics” and towards more specific, desired styles.

The overarching lesson from the Opus 5.5 migration guide is that while model capabilities are rapidly advancing, the frameworks and application logic supporting them often lag. Successfully deploying AI agents now requires a nuanced understanding of how to manage AI communication, define task completion, maintain context, and allocate computational resources effectively. The responsibility for an agent’s success or failure is split between the AI’s power and the developer’s implementation.

Source: https://www.ithome.com/1/007/442.htm