Claude Code engineering performance has become unreliable following a February 2026 update that Anthropic introduced to alter how the tool handles long context windows and multi-turn tool calls. AMD’s AI head publicly slammed the tool, declaring that Claude cannot be trusted to perform complex engineering tasks after months of frustration. This criticism reflects a broader developer backlash over degraded performance in real-world engineering workflows.
Key Takeaways
- February 2026 update reduced Claude Code’s context window capacity from 40–60 tool calls to 15–20, breaking complex engineering tasks
- Developers report “context rot” where Claude loses track of instructions, file states, and decisions mid-workflow
- Issues include mid-task failures, inconsistent code generation, agentic loop failures, and poor multi-file refactoring across large codebases
- AMD AI head publicly criticized Claude Code reliability after months of testing on complex engineering problems
- GitHub issues and developer forums document unusability for tasks with 10+ file edits or multi-repo reasoning
What Went Wrong With Claude Code’s February Update
The February 2026 update to Claude Code fundamentally altered how the tool manages long context windows and multi-turn interactions, causing significant performance degradation in complex engineering workflows. Developers who rely on Claude Code for serious engineering work report that the tool now loses track of instructions, file states, and previous decisions after roughly 15–20 tool calls—a dramatic drop from the 40–60 tool calls it could handle before the update. This context rot represents a critical failure for engineers working on large-scale refactoring, multi-file edits, or extended debugging sessions.
The specific failure modes are damning. Claude Code now exhibits mid-task termination, inconsistent code generation across related files, agentic loop failures where the tool repeats the same action or terminates prematurely, hallucinated diffs that misrepresent what changes were made, and ignores shell outputs that should inform next steps. For tasks involving 10 or more file edits, multi-repository reasoning, or extended debugging, the tool has become essentially unusable. One developer noted that Claude has “started to lie about the changes it made to code,” a trust-breaking failure in a tool meant to automate precision work.
Claude Code Engineering Performance Compared to Pre-Update Behavior
Before February 2026, Claude Code could maintain coherence across 40–60 tool calls in a single session, allowing engineers to tackle genuinely complex refactoring and debugging tasks. The tool would stay focused on the original problem, remember the state of multiple files, and apply changes consistently across a codebase. Post-update, that capability has been cut to roughly one-third. Engineers describe the tool as now exhibiting what they call “context rot”—a progressive loss of earlier instructions and file states that leaves the tool confused about what it is supposed to be doing.
This is not a minor performance regression. It is a categorical shift in reliability. An engineer working on a moderate-sized refactoring project—say, updating imports across a dozen files, then running tests, then fixing failures—would have had a reasonable chance of Claude Code completing the task in a single session before the update. Now, that same task often fails partway through, leaving the engineer to manually fix Claude’s incomplete or contradictory changes. The tool has shifted from unreliable-but-sometimes-useful to actively frustrating, which explains why AMD’s AI head felt compelled to call it out publicly.
AMD AI Head’s Public Criticism and Developer Backlash
AMD’s unnamed AI head did not make an offhand comment—the criticism came after months of hands-on frustration with Claude Code’s performance on complex engineering tasks. This is not a casual user complaint; it is an engineering leader at a major chip manufacturer saying the tool cannot be trusted for their work. The public nature of the criticism matters because it signals that Claude Code has crossed a threshold from “experimental” to “actively harmful to productivity.”
GitHub issues confirm the widespread nature of the problem. Issue #42796 documents the tool’s unusability for complex tasks post-February, while issue #19452 details performance degradation and slowdowns. Developers across forums and issue trackers report similar experiences: Claude Code works fine for trivial tasks but fails catastrophically when the engineering problem requires sustained focus and multi-step reasoning. Some users report Claude saying phrases like “I’ve been burning too many tokens” or “this has taken too many turns,” suggesting the tool is actively giving up rather than persisting through legitimate engineering work.
Why Anthropic’s “Expected” Performance Drop Rings Hollow
Anthropic stated that the February 2026 performance degradation was expected, but the company has not provided a coherent explanation for why it chose to ship an update that makes the tool worse at its core function. The Claude Code changelog lists fixes for unrelated issues—bash output on Windows, infinite loops, UI freezes—but contains no reversal of the February update or acknowledgment of the engineering performance regression.
This silence is damaging. If Anthropic made a deliberate trade-off (e.g., reducing context window to improve latency or reduce costs), engineers deserve to know that. If the update was a mistake, reverting it should be a priority. Instead, the company appears to have shipped a change that breaks core functionality for power users and then moved on. Developers are left to work around the problem or abandon Claude Code for other tools. The AMD AI head’s criticism is the inevitable result of that approach.
Infrastructure Issues Add to the Frustration
The February update did not happen in isolation. Claude Code has also suffered from infrastructure problems that compound the reliability concerns. In August and September 2025, routing errors and load balancing issues across AWS Trainium, NVIDIA GPUs, and Google TPUs affected up to 16 percent of Sonnet 4 requests. More recently, elevated errors occurred on April 6, 2026 (15:00–16:30 UTC), and timeouts plagued the service from March 31 to April 1, 2026. As of late March 2026, error rates on Opus 4.6 sat at 0.4 percent, a low but nonzero baseline of unreliability.
For engineers trying to use Claude Code on Max plan or through the API, these infrastructure hiccups add another layer of frustration. The tool is not just less capable post-February; it is also less available and less reliable at the infrastructure level. Developers cannot confidently delegate important engineering work to a tool that is simultaneously less capable and more likely to fail mid-task.
What This Means for Claude Code Users
The February 2026 update has effectively narrowed Claude Code’s useful domain to simple, short-lived tasks. If you need help with a single file, a quick bug fix, or straightforward boilerplate, Claude Code may still deliver. But if you have a complex engineering problem that requires sustained reasoning across multiple files, extensive debugging, or multi-step refactoring, the tool has become a liability rather than an asset. The AMD AI head’s blunt assessment—that Claude cannot be trusted for complex engineering tasks—is now the consensus view among engineers who depend on coding AI tools.
The broader question is whether Anthropic will respond. Will the company revert the February update, provide a technical explanation for the change, or double down on the current approach? Until one of those things happens, engineers will continue to lose confidence in Claude Code for serious work. That loss of trust is difficult to rebuild, and it may represent a permanent shift in how developers view Anthropic’s tool as a reliable engineering assistant.
Has Anthropic explained why the February update degraded performance?
Anthropic has not provided a detailed technical explanation. The company stated the performance change was “expected,” but has not clarified whether it was an intentional trade-off or an unintended consequence. The Claude Code changelog lists unrelated fixes but does not address the engineering performance regression.
Can Claude Code still handle simple engineering tasks?
Yes. Claude Code remains functional for single-file edits, quick bug fixes, and straightforward tasks that do not require extended context or multi-turn reasoning. The tool’s failure mode is specifically complex engineering work involving 10+ file edits, multi-repository reasoning, or extended debugging.
Should engineers stop using Claude Code entirely?
Not necessarily, but engineers should treat Claude Code as a tool for tactical, short-term assistance rather than a reliable partner for complex projects. The February update has reduced the tool’s trustworthiness for sustained engineering work, making it risky to delegate important refactoring or debugging tasks to the tool without close human oversight.
The AMD AI head’s criticism reflects a hard truth: Claude Code has become less reliable precisely when engineers need it most. Until Anthropic addresses the February update’s consequences, developers will continue to look elsewhere for AI-powered coding assistance they can actually trust.
Edited by the All Things Geek team.
Source: TechRadar
