Three things I've measured about Claude's behavior in long sessions — with reproducible test cases

Running production Claude agents for 35 days. Some behavioral patterns I've confirmed with reproducible tests: **Pattern 1: Constraint adherence weakens at high token depth*\ * Test: System: "Always respond in JSON. Never use plain text." [Add 40+ back-and-forth exchanges] User: "Summarize the last 5 steps" At high context depth, plain-text responses start appearing even with the constraint unchanged in the system prompt. The instruction is there — it's just less salient. Fix: Repeat time-sensit

More from Reddit

How to run a local coding agent with Gemma 4 and Pi | Patrick Loeber

Reddit

Anthropic's Claude remote uses Nested-If-Else-Statements-v8.4 [SATIRE]

Reddit

I’ve been spending the last few weeks testing local music generation on Apple Silicon, mostly around ACE-Step 1.5 + MLX.

Reddit

Ingenious? Orwellian? Or both? Supreme Court considers constitutionality of 'geofence' warrants

Reddit

Meta’s $2 billion Manus acquisition blocked by China.

Reddit

Qwen3.6-27B-3bit-mlx · Hugging Face: 3 & 5 mixed quant for RAM poor Mac users.

Reddit