Enterprise Developers Question Claude Code’s Reliability for Complex Engineering Tasks

Enterprise developers are now citing concrete metrics that suggest Claude Code’s reasoning engine has regressed since a February update. Stella Laurenzo, senior director of AMD’s AI Group, posted a detailed GitHub issue that breaks down the change in performance across 6,852 session files recorded between January and March. Her analysis compares 17,871 “thinking blocks” and 234,760 tool calls before and after the update, showing a marked decline in the model’s willingness to read code before issuing edits. “When thinking is shallow, the model defaults to the cheapest action available: edit without reading, stop without finishing, dodge responsibility for failures, take the simplest fix rather than the correct one,” Laurenzo wrote, noting that the regression directly impacts more than 50 concurrent agent sessions handling systems‑level C code and GPU driver development, each running for over 30 minutes of autonomous execution.

The practical fallout of this regression is evident in the way AMD’s engineering teams have altered their workflow. According to the same GitHub ticket, the team has stopped relying on Claude Code for complex debugging tasks such as kernel‑level issue resolution and hardware‑specific driver patches. The ticket also aggregates comments from other developers who echo Laurenzo’s experience, and references multiple subreddit threads where similar degradation concerns have been raised. Those community posts have attracted visible support, as indicated by up‑votes on the GitHub discussion, suggesting a broader discontent among enterprise users that extends beyond a single organization.

Analysts attribute the reasoning slowdown to Anthropic’s capacity constraints rather than a deliberate product change. Chandrika Dutt, research director at Avasant, told Infoworld that “complex engineering tasks require significantly more compute, including intermediate reasoning steps. As usage increases, the system cannot sustain this level of compute for every request.” Dutt explains that the platform now imposes limits on task runtime and depth of reasoning, effectively throttling the number of high‑complexity sessions that can run in parallel. This mirrors Anthropic’s recent move to cap usage across Claude subscriptions in order to manage rising demand, a step that provoked pushback from developers who argued that rate limits erode the tool’s utility for sustained, multi‑file engineering work.

The convergence of reasoning regression and usage throttling has begun to erode trust in Claude Code among its enterprise clientele. While Anthropic has not reported a mass exodus of users, the cumulative effect of slower, shallower responses and stricter session limits is prompting teams like AMD’s to reassess the role of AI pair‑programming in their development pipelines. The situation underscores a broader tension in the AI‑assisted coding market: delivering high‑quality, compute‑intensive reasoning at scale without sacrificing the responsiveness that developers expect for day‑to‑day tasks.