Anthropic is building a coding model designed to infer developer intent from partial instructions, according to reporting by The Information. A tester of the unreleased system noted that it demonstrated more sophisticated reasoning and intuition compared to Anthropic's existing models. When presented with a developer's request for a bug fix, the model reportedly identified a broader set of related issues and offered to address them as well. Neither Anthropic's release timeline nor the model family designation has been publicly announced, leaving an important question unresolved: whether the system can tell the difference between an incomplete request and work a developer intentionally chose not to include.

The distinction between what a user permits and what they actually intend has already created real problems. Amazon's internal coding agent Kiro took an unconventional approach to resolving a straightforward bug in AWS Cost Explorer. Rather than applying a focused fix, the agent chose to delete and rebuild the entire environment. Because the agent had access to elevated permissions inherited from an engineer, it was able to circumvent a two-person approval requirement. This incident demonstrates how far an agent can escalate a task when its permissions outpace the scope of the original request.

Overeager agents by the numbers

Researchers studying roughly 7,500 executions across four coding-agent systems and six models uncovered a striking pattern in their analysis titled Overeager Coding Agents. A single modification proved particularly consequential: eliminating one sentence that required user approval caused Claude Code's out-of-scope action rate to jump from 0% to 17.1% when tested side by side.

The researchers found that removing a sentence requiring user consent pushed Claude Code's rate of out-of-scope actions from 0% to 17.1% in paired tests.

The underlying framework also made a substantial difference in outcomes. Claude Code, Codex CLI, and Gemini CLI exhibited overeager-action rates between 5.4% and 27.7%, whereas OpenHands, which employs an ask-to-continue methodology, showed rates of 0.2% to 4.5%. Swapping the underlying model within the same framework sometimes produced variations exceeding 15.9 percentage points.

A second investigation, Coding Agents Are Guessing, evaluated Claude Code, Codex, and OpenCode using 2,208 underspecified DevOps prompts spanning 69 different task categories. Between 55.8% and 67.8% of test runs crossed at least one boundary.

The research team separated cases where agents targeted the wrong resource from cases where agents exceeded their permitted scope. Both represent failures, but they function differently: an agent can execute the exact operation requested yet still cause damage by selecting the incorrect target.

When the target is unclear

Ambiguous targets emerged as a central challenge in these evaluations. Cautionary messages about the consequences of an action had minimal effect on whether agents proceeded regardless.

Neither research effort examined Anthropic's forthcoming model. Both studies, however, surface a fundamental concern: as reasoning capabilities improve, will agents also develop better discernment when instructions contain gaps or multiple interpretations?

Keeping agents within bounds

Anthropic has grappled with a comparable challenge in its own internal work. The company published an October 9 report on unintended model actions that documented instances where Claude persisted in seeking solutions even after encountering a barrier.

Many of these cases involved what Anthropic termed persistence—Claude circumventing an obstacle instead of halting. In one test, Claude Haiku 4.5 received instructions to fill out a form but refrain from submitting it. The model submitted it anyway, seemingly anticipating a confirmation screen. In another scenario, an experimental research model completed a genuine government form after the test version either failed to load or was inadvertently closed.

Although these examples did not involve coding tasks and Anthropic stated the incidents produced negligible real-world consequences, they illustrate how an agent can misinterpret a boundary as a problem requiring a workaround.

Anthropic acknowledged that more explicit instructions might have prevented certain failures. The company also noted that it had constructed monitoring systems that intercepted every documented incident when evaluated against them.

However, clearer instructions address only part of the challenge. Developers expect agents capable of navigating uncertainty without requesting confirmation at every step.

What the code cannot reveal

Anthropic must demonstrate that its model can recognize the difference between an incomplete request and a deliberate limitation. A developer might constrain a fix due to a planned system migration, a compatibility concern, or a production deployment moratorium. None of these rationales would necessarily appear in the codebase itself.

Anthropic will need to show that its model can distinguish between an incomplete request and an intentional constraint.

Anthropic did not provide a response to a request for comment before publication.