Stale Instructions Override Code in 90% of Agent Runs, Plus Background MCP Tasks and Cache-Spend Audits

Today's findings

1. #1 Claude Code 2.1.319 Adds Non-Prompting Permission Denials and Background MCP Taskstool

   Claude Code can now deny prompt-waiting tool calls instead of stalling, while MCP tasks move to the background and loop semantics tighten for unattended runs.

   Claude Code 2.1.319

   Permission prompts no longer stall unattended runs

   Before · wait on prompt

   • Tool call hits permission wall
   • Headless session hangs
   • CI & subagents wait idle
   • MCP tasks block up front

   → Hang points removed; env vars restore the old sync behavior

   After · deny & continue

   • Call denied, run proceeds
   • Sessions run to completion
   • Unattended runs keep flowing
   • MCP tasks go background

   Escape hatches: CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND or CLAUDE_CODE_DISABLE_MCP_TASKS keep MCP synchronous.

   Why it matters: Small teams running long or headless Claude Code sessions need predictable permission behavior; this update removes common hang points and gives env-var escape hatches when integrating MCP.

   How to apply: Upgrade to 2.1.319, test permissions.denyInsteadOfPrompting in CI/subagent runs, and set CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND or CLAUDE_CODE_DISABLE_MCP_TASKS if your workflow needs synchronous MCP behavior.

   claude-codeagentsmcppermissions

   Read more: Claude Code 2.1.319: new denyInsteadOfPrompting setting, MCP tasks go to the background, /loop behavior change

2. #2 Stale CLAUDE.md Files Override Code in 90% of Agent Runspaper

   A 320-run study found agents often notice stale instruction files but follow them anyway; one user-level instruction to trust code over the file raised code-following from 4% to 91%.

   Why it matters: Agent instruction files are configuration, not just docs—stale rules silently cause wrong edits. A single precedence rule can prevent expensive drift.

   How to apply: Add a user-level instruction: when an instruction file conflicts with the code, trust the code and say so. Audit skills and CLAUDE.md for dead paths, old model names, and unused entries.

   agentsclaude-codeinstructionsevaluation

   Read more: A study found agents follow a stale CLAUDE.md 90% of the time, even when they notice the conflict

3. #3 Llama-modchip Runs 284B MoE Models on 64GB RAM Without a GPUrepo

   An open llama.cpp-based project places expert weights across VRAM, RAM, and flash so MoE models larger than available memory can still prefill and decode.

   Llama-modchip

   How a 284B MoE fits in 64GB

   HOT

   VRAM Most-routed experts

   WARM

   System RAM Warm expert cache

   COLD

   NVMe flash Rare experts, streamed

   Expert weights tiered by routing frequency across VRAM, RAM, and flash — 64GB RAM, no GPU required.

   Why it matters: This makes larger open MoE models testable on laptops, phones, and modest GPUs instead of assuming server-class memory; it also clarifies when SSD offload is practical.

   How to apply: Try llama-modchip or the related llama.cpp placement tooling on a small MoE first, watch expert-cache hit rates and NVMe read latency, and keep frequently routed experts in RAM/VRAM.

   local-llmllama-cppmoeoffloading

   Read more: Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU · Run MoE models bigger than your RAM on phone, laptop or GPU: it places every weight on VRAM, RAM or flash by itself for max prefill and decode speed (built on llama.cpp)

4. #4 Claude Code Cost Hinges on Model Choice and Cache Readstip

   A 984-run commit benchmark found Haiku about 6× cheaper per solved task than Sonnet, while a separate session audit found cache reads were 66% of spend.

   Agent cost economics

   Cost per solved task, not sticker price

   6×

   cheaper per solved task — Haiku vs Sonnet

   984-run commit benchmark

   66%

   of session spend was cache reads

   2

   cost levers: model choice + cache reuse

   Well-specified tasks suit Haiku; audit cached prefixes before scaling up to Opus.

   Why it matters: Agent costs are dominated by context reuse and failed attempts, not just output tokens. Measuring cost per solved task and cache behavior is more useful than sticker price.

   How to apply: Log response IDs to avoid double-counting, compute cost per passing test/task, try Haiku for well-specified work, and reduce or restructure large cached prefixes before scaling up to Opus.

   claude-codecostbenchmarkscaching

   Read more: I reran 56 real commits with Haiku, Sonnet and Opus 5.5 in Claude Code (984 runs). Here's the cost per solved task · I priced 194 of my Claude Code sessions at Opus 5.5 API rates. 66% of the cost is cache reads, not output

5. #5 Production RAG Lessons: Keep Simple RAG, Add Agents Only When Accuracy Paystechnique

   A year of healthcare RAG work found classic retrieval is still best for latency/cost, while agentic RAG raised cost 2–3× for 25–30% accuracy gains.

   Production RAG lessons

   Agentic RAG: +25–30% accuracy for 2–3× cost

   vs

   Classic RAG

   Agentic RAG

   Cost

   1× baseline

   2–3× per call

   Latency

   Fastest

   Agent loops add delay

   Accuracy

   Baseline

   +25–30%

   Classic RAG wins the row Agentic RAG wins the row

   Keep simple RAG as the default; spend agent loops only where the accuracy lift pays.

   Why it matters: It gives a concrete tradeoff for teams deciding whether to add agent loops to RAG instead of defaulting to complexity.

   How to apply: Start with optimized chunking, retrieval, and reranking; instrument accuracy and cost; reserve iterative/agentic retrieval for queries where the accuracy lift justifies the extra calls.

   ragagentsretrievalproduction

   Read more: What I've learned about RAG and Agentic RAG after one year in enterprise production

6. #6 Tool Descriptions Are Safety Rails for Local Support Agentstip

   A fully local Ollama+LangChain support agent refunded three customers instead of one after a vague refund-tool description, changing behavior with the same model and prompt.

   Why it matters: Tool schemas and descriptions are executable policy for agents; small wording differences can create real financial or data side effects.

   How to apply: Write narrow tool descriptions with explicit preconditions and limits, test adversarial cases before enabling write actions, and gate refunds or spend behind deterministic validation or human approval.

   agentslocal-llmollamatool-use

   Read more: I built a fully local AI support agent (Ollama + LangChain) and one bad tool description made it refund 3 customers instead of 1

7. #7 Caminus Brings MoE Inference to Dual AMD Radeon GPUstool

   Caminus 0.2.0 is an open LLM inference engine for AMD GPUs with custom HIP kernels, multi-GPU expert parallelism, and one kernel launch per token.

   Why it matters: It gives local-inference teams an AMD/ROCm path for MoE models beyond CUDA-only stacks, where tooling is still sparse.

   How to apply: Test Caminus on supported RDNA3 cards, benchmark prefill/decode and multi-GPU scaling against llama.cpp or ROCm baselines, and inspect HIP kernel requirements before production use.

   local-llmamdrocmmoe

   Read more: Caminus: MoE inference on two AMD Radeon GPUs (RDNA3, ROCm/HIP), one kernel launch per token.

8. #8 35 Domain-Expert LoRAs on Qwen3.6 MoE, Trained and Served on a Mac Studiorepo

   An Apache-2.0 release shares 35 Qwen3.6-35B-A3B LoRA adapters, datasets, configs, router, and scripts; BF16 runtime adapter hot-swapping worked while fusing into 8-bit failed.

   architecture

   One MoE base, 35 domain experts — one hot-swapped per request

   router

   top-1 gate

   expert LoRAshot-swapped per request

   weighted

   merge

   35

   Qwen3.6-35B-A3B adapters, Apache-2.0

   BF16

   runtime hot-swap worked

   8-bit

   fused adaptation failed

   Mac Studio

   trained and served on one machine

   Why it matters: It is a practical blueprint for local multi-domain specialization on MoE models, including serving and routing failure modes.

   How to apply: Use the published configs as a baseline, serve one base model with BF16 adapters swapped per request, and treat routing as the main bottleneck to evaluate before adding more domains.

   loramoemlxfine-tuning

   Read more: 35 domain-expert LoRAs on a MoE (Qwen3.6-35B-A3B) trained with MLX on one Mac Studio — what worked, what didn't, all Apache-2.0

9. #9 OXR 1.3B Document Parser Trains on Its Own Extraction Mistakesrepo

   A small open document model splits layout from recognition and uses error-sorted post-training to output reliable tables, formulas, Markdown/JSON, and bounding boxes.

   OXR 1.3B document parser

   A parser that trains on its own mistakes

   1

   Parse PDFs

   tables, formulas, boxes

   2

   Rank errors

   confidence sorts worst extractions

   3

   Retrain

   mistakes become training data

   each pass, fewer mistakes

   Open 1.3B weights: cheaper and more debuggable than giant multimodal parsing calls.

   Why it matters: PDF table extraction is a recurring RAG/data problem; a small traceable parser can be cheaper and more debuggable than giant multimodal calls.

   How to apply: Try OXR on merged-header and multi-page-table PDFs, keep layout analysis swappable, and use confidence/uncertainty sorting to retrain on the model's own bad extractions.

   ragdocument-parsingpost-trainingpdf

   Read more: What we learned post-training a 1.3B document parser on its own mistakes · What we changed in a small document parser to get tables out of PDFs reliably

10. #10 Curated Claude Plugins, Skills, and MCP Servers Organized by Taskrepo

    A repo indexes Claude plugins, skills, and MCP servers by intended job—video editing, SEO, design, dev, and more—with one-line descriptions.

    Curated Claude toolbox

    Shelved by the job to be done — not dumped in a pile

    Video Editing & motion

    SEO Search optimization

    Design Visual work

    Dev Code & tooling

    More Lanes keep growing

    Curated task lanes '…and more'

    Browse by task to cut discovery time — and review MCP permissions before installing into shared environments.

    Why it matters: It cuts discovery time when wiring Claude Code into real workflows and helps teams avoid rebuilding common MCP integrations.

    How to apply: Browse by task, vendor or pin useful repos, and review permissions/security before installing MCP servers or plugins into shared agent environments.

    claudemcppluginstools

    Read more: I made a repo of Claude plugins and skills, organised by what you want to do

11. #11 Weight-Space Programming Harness Lets Opus Edit Another Model's Weightspaper

    An open test harness applies the MIT CSAIL weight-space programming idea, using Claude Code to make small direct weight edits without retraining.

    Why it matters: It turns model editing into an agentic debugging task and may offer targeted fact removal or repair, but it needs rigorous validation before production.

    How to apply: Use the harness on a disposable open model, define a narrow behavior/fact target, compare before/after evals, and keep edits versioned and reversible.

    model-editingclaude-codepaperopen-source

    Read more: Weight Space Programing with Opus - Manipulating weights natively using Opus to make a model forget facts it learned