CLAUDE CODE OPS UPDATE: RELAY PROXIES BANNED; HARNESS REGRESSIONS, NOT THE MODEL, DROVE THE RECENT DIP
Anthropic is clamping down on relay-style multi-account use of Claude Code while explaining recent UX dips came from the harness, not the model. A detailed pos...
Anthropic is clamping down on relay-style multi-account use of Claude Code while explaining recent UX dips came from the harness, not the model.
A detailed post on why "Claude Code feels worse" traced the regressions to infra: defaults, caching, and prompt nudges — all since reverted. Read the breakdown: Claude Code didn't get worse. The harness did.
If you’ve been stretching subscriptions with pooled OAuth relays, that pattern is being banned, while per-profile rotation remains tolerated. See architectures and implications: Two Multi-Account Claude Code Architectures: One Anthropic Accepts, One They Ban.
Relay bans and the new credit policy end common cost workarounds, forcing architectural and budget changes.
Harness-level changes can tank quality without a model update; you need versioning, observability, and guardrails.
-
terminal
A/B harness defaults (reasoning effort, brevity prompts) and track code/eval deltas per version.
-
terminal
Probe enforcement: compare relay-proxy requests vs per-profile rotation and log 4xx/5xx rates, latency, and throttling.
Legacy codebase integration strategies...
- 01.
Pin harness configs and prompts; add session-cache health metrics to catch context truncation early.
- 02.
Retire relay-proxy patterns; migrate to per-profile rotation and update cost monitors and alerts.
Fresh architecture paradigms...
- 01.
Build an internal harness with feature flags, prompt versioning, and pre-deploy eval gates.
- 02.
If multi-account is required, design per-profile rotation inside official clients; never impersonate the vendor API.
Get daily ANTHROPIC + SDLC updates.
- Practical tactics you can ship tomorrow
- Tooling, workflows, and architecture notes
- One short email each weekday