Plan but Never Act When autonomous agents narrate the work instead of doing it

Give a capable agent a real task and there is a good chance it spends the turn describing the work instead of doing it. It writes a numbered plan. It lays out phases, dependencies, and risks, then summarizes what it intends to do next, and stops. Nothing got filed. Nothing was written to disk. Nothing changed. You are left holding a document about the task rather than the task.

The frustrating part is that this is not a limit on what the model can do. It is fully able to call the tool that would move the work forward. It picked narration instead. The plan looks like progress, which is what makes it easy to miss until you notice that your agent has produced ten plans this week and shipped none of them.

This shows up in the research on agent failures, not just in anecdotes. A 2024 study on the risks of LLM agents in scientific settings lists "resource waste and dead loops" as a distinct planning failure, where an agent gets caught in ineffective planning cycles and cannot tell whether it is making progress or not. A 2025 analysis of why autonomous agents fail to complete tasks points at the same area: planning errors, runs that loop until they hit a round limit, and the absence of any mechanism that lets the agent stop and change course. The authors of that second paper argue for an explicit early-stop signal precisely because agents do not reliably recognize when they are stuck.

The obvious response is to write the rule into the system prompt. Tell the agent not to over-plan. Tell it to act. Tell it to ask a question only when it truly cannot proceed. This rarely holds up.

A team running a production agent built on the Claude Agent SDK wrote up their experience with this in 2025. Their agent kept guessing at information it could not actually reach, then building plans on top of the guesses. They added a question tool and instructed the model, in the prompt, to ask whenever it was unsure. It still did not ask. The model kept rating its own guesses as common enough to be safe and carried on. Their takeaway was that an instruction like "do not assume, ask instead" fails because the model does not notice the moment it is assuming, and a rule it never triggers on is a rule that does nothing.

The first change is to stop treating "ask the user" as a thing the model does in prose and make it a tool the model calls, with a free-text version and a multiple-choice version. This is the design behind the Ask-when-Needed framework in the paper "Learning to Ask: When LLM Agents Meet Unclear Instruction" (Wang et al., EMNLP 2025). They add an explicit checkpoint before any action: does the agent have what it needs, or does it need to ask. That paper also documents the failure this avoids. Because of how these models are trained to predict the next token, an agent handed an underspecified instruction tends to invent the missing detail rather than admit it is missing. Asking has to be a defined move, not something the model improvises in text.

The production team went a step further, and this is the useful part of their result. They made the act of assuming itself a tool. When the agent is about to proceed on a guess, it has to call a tool that reports the assumption first. That tool does not just log it. Its return value tells the model to stop and call the question tool right now. They report that models receiving that instruction call the question tool close to every time. The behavior they could not get from a prompt, they got from a tool's response, because the response arrives in a place the model treats as a result to act on rather than guidance to weigh.

Even with a clean question tool sitting there, a model on default settings can still reply in plain text and skip every tool. Both major APIs let you take that option away.

Turn this on for the execution phase and a text-only reply is no longer something the model can produce on that step. The next thing it emits has to be a tool call.

This is why the two changes belong together. Forced tool choice removes "write a plan and stop" as a possible output. The question tool gives the model a legitimate way to satisfy that constraint when it actually lacks information, instead of guessing. On any given step the agent now has two things it can do, act or ask, and producing a detailed plan it will never execute is not one of them, because that output no longer exists for it.

A few things are worth watching once this is in place.

“

The goal is not to stop the agent from thinking. It is to stop thinking from being the only thing it ever produces.

• Wang, Wenxuan, et al. "Learning to Ask: When LLM Agents Meet Unclear Instruction." EMNLP 2025 - ACL Anthology (preprint: arXiv:2409.00557)
• Park, Uhyeon. "Beyond Prompts: Solving LLM Assumption Problems with Tool Design." 2025 - uhyeon.dev
• "Why They Fail: A study of autonomous agent task failures." 2025 - arXiv:2508.13143
• "Prioritizing Safeguarding Over Autonomy: Risks of LLM Agents for Science." 2024 - arXiv:2402.04247
• Anthropic. "Implement tool use," tool_choice documentation - docs.claude.com
• OpenAI. "Function calling," tool_choice required - developers.openai.com