RELIABILITY

CLAUDE_CODE_RETRY_WATCHDOG=1 makes Claude Code wait out 429 and 529 errors instead of failing after 10 retries. When to use it, how to cap it, what we measured.

The August 17 GitHub outage lasted 7 hours 47 minutes and disrupted PRs, Actions, APIs, and Copilot. If your coding agents treat GitHub as the control plane, you need a local fallback that can keep shipping for four hours.

A new arXiv paper probes six widely used open-source agent frameworks and finds the barrier semantics of approval gates, cancellation, and timeouts hold on none of them. A sibling branch can execute while the user is rejecting another one, and replay can double-execute. The fix is a verified external gate called SoundGate.

A new ASE 2026 framework injects server errors, truncated responses, and corrupted tool calls into live agent systems at the HTTP layer. Every system degrades, pass@1 drops up to 50 points, and the ranking stays the same no matter which LLM is behind it.

A new arXiv paper builds a deterministic monitor on top of SWE-agent that watches long agent trajectories and only calls an advisor LLM when the run actually drifts. Resolution rates go up by up to 15.2 points at an extra $0.08 per instance, and the paper argues the expensive approach is re-planning from inside the loop.

Fable 5 refusals come back as a 200 response, not an error. At fleet scale, that quietly corrupts entire runs. Here is how to detect, fall back, and treat refusal rate as a health metric.

Claude outages and 529 overloads expose whether your AI coding workflow has checkpoints, receipts, model-switch paths, and small enough task slices to survive provider degradation.

LangChain's rubrics for Deep Agents point at a practical agent pattern: self-correction works only when rubrics are versioned, executable, and sampled against human review.

A long-running coding agent is only useful if the environment around it can queue tasks, capture logs, checkpoint state, verify behavior, limit cost, and recover from failure.

The defensive patterns that keep Claude integrations alive in production. Retry shapes, backoff with jitter, circuit breakers, fallback chains, and the observability you need to debug at 3am.

The math of agent pipelines is brutal. 85% reliability per step compounds to about 20% at 10 steps. Here is why long chains collapse in production, and the six patterns the field has converged on to fight the decay.