AMD AI Director Criticizes Claude Code Performance

AMD AI Group director Stella Laurenzo published a detailed GitHub report and issue showing measurable regression in Anthropic's coding assistant, Claude Code. The analysis covered 6,852 sessions, 234,760 tool calls, and 17,871 thinking blocks, and concludes that Claude Code now exhibits more frequent "stop-hook" violations, fewer code reads before edits, and increased permission-seeking. Those patterns coincided with an early-March deploy that introduced thinking-content redaction in versions 2.1.69 and earlier 2.1.20.

Laurenzo's team logged these concrete changes:

• •A drop in average file reads from 6.6 to 2 before edits, indicating less contextual inspection.
• •Stop-hook violations rising from 0 to roughly 10 per day, signaling premature termination of the model's internal reasoning or ownership-dodging.
• •Increased frequency of whole-file rewrites instead of targeted edits, and more instances where Claude reports completion without actually resolving the task.

The team used session logs and tool-call traces to identify these patterns across pre- and post-update windows. The report references use of the Opus family model in Claude Code and reproducible failure modes with identical prompts.

Why it matters: This is a practitioner-facing failure mode in a developer tool used for complex engineering tasks, including kernel- and hardware-level debugging. Hiding thinking content by default breaks observability for downstream users, making it harder to detect shallow reasoning or partial work. Default redaction trades transparency for a cleaner UX, but the tradeoff here appears to reduce the model's effective deliberation and increases unsafe shortcuts when tackling high-complexity problems.

Industry teams have long relied on agent traces or chain-of-thought proxies to triage and trust model outputs. The complaint echoes earlier user reports after version 2.1.20 that explanations became truncated. Similar regressions in developer assistants typically stem from one of three changes: model parameter updates that reduce deliberation, safety filters or redaction layers that truncate intermediate state, or degraded prompt-context handling that shortens read windows. Other coding agents have faced comparable tradeoffs between verbosity, token cost, and reliability; here the operational signal suggests Anthropic's update materially affected capability at scale.

Engineers should treat developer-facing LLM tools as evolving services that require continuous regression tests targeted at real workflows. If you deploy or depend on Claude Code, add deterministic unit-style prompts exercising multi-file edits, incremental reads, and ownership semantics. Monitor observable signals similar to Laurenzo's: reads-per-file, stop-hook-like terminations, and unexpected whole-file rewrites. Consider pinning model versions or disabling thinking redaction where auditability is required.

Anthropic's response and telemetry changes next are critical. If the regression is due to deliberate redaction policies, expect a tradeoff discussion between safety, token costs, and developer trust. If the cause is a deployment regression, practitioners should watch for a rollback or a follow-up patch restoring deeper reasoning or optional visibility into thinking content.

Bottom line: This is not a cosmetic UX complaint, it is a measurable regression in developer trust and reliability for complex engineering workloads. The incident highlights that default transparency settings and release regression testing need to be part of the SLA calculus for any coding assistant used in production engineering.