Blue Moose logo Blue Moose Field notes
Tag / AI
Agent Skills are just folders of Markdown — and after a year of MCP hype, that's the point
ai

Agent Skills are just folders of Markdown — and after a year of MCP hype, that's the point

The most useful upgrade to my agent workflow lately is embarrassingly low-tech: a folder with a SKILL.md in it. Anthropic's Skills (now an open standard) use progressive disclosure to load the right instructions only when relevant. They're the honest counterweight to MCP: half the time you don't need a live tool, you need the paragraph the agent was missing — versioned in git, no daemon. Plus the supply-chain caveat nobody screenshots.

BM Blue Moose / 2026-09-07 / 04 min
2026-09-07 / 04 min
RAG cost per query is a lie — the real bill is the re-index you haven't done yet
ai

RAG cost per query is a lie — the real bill is the re-index you haven't done yet

Everyone budgets RAG by the per-query LLM call, then gets blindsided three sprints in when they change chunking or swap embedding models and have to re-embed the whole corpus. Real numbers, why an embedding model is a lock-in decision disguised as a benchmark score (incompatible vector spaces), and how to budget for the re-index — plus when it genuinely doesn't matter.

BM Blue Moose / 2026-09-02 / 04 min
2026-09-02 / 04 min
WebMCP: your website becomes a tool the browser agent calls — read the fine print
ai

WebMCP: your website becomes a tool the browser agent calls — read the fine print

A Google/Microsoft-backed proposal lets your page hand an AI agent real callable tools instead of screenshot-and-guess. Chrome's already trialing it. But it's a W3C draft (not a standard), the API is churning, and the security fine print is stark: tools run with your logged-in auth and there's no mandated consent model. Clever, and a new attack surface — both at once.

BM Blue Moose / 2026-08-24 / 04 min
2026-08-24 / 04 min
The open-weight coding gap closed — so why are your agents still worse?
ai

The open-weight coding gap closed — so why are your agents still worse?

GLM-5 (MIT), Kimi K2.6, DeepSeek V4 and Qwen3.6 all caught up to the frontier this year. I swapped my agent to a top open-weight model expecting better diffs and got the same occasional failures — because the model was never the bottleneck. The harness is (tools, context assembly, retries, verification). Plus the self-hosting asterisk: 'open weights' ≠ 'runs on your box'.

BM Blue Moose / 2026-08-24 / 04 min
2026-08-24 / 04 min
It deleted prod in 9 seconds — and the AI was the least interesting part
ai

It deleted prod in 9 seconds — and the AI was the least interesting part

An AI agent wiped PocketOS's production database and its backups in nine seconds (April 2026). Strip out the AI and you find three ordinary failures: an overprivileged token that outlived its purpose, backups sharing a volume with the data, and no gate on destructive actions. Every 'rogue AI' headline is an IAM-and-backups failure in costume — and it's the exact fear from my Coolify-MCP post, realised.

BM Blue Moose / 2026-08-17 / 04 min
2026-08-17 / 04 min
The junior-dev cliff is real — but it's a training-pipeline problem, not a layoff story
ai

The junior-dev cliff is real — but it's a training-pipeline problem, not a layoff story

Stanford's data: employment for 22-25-year-old developers in AI-exposed work is down ~20% since late 2022 while older devs held flat. The scary chart is real; the wrong lesson is 'AI took the jobs.' The slower, worse one: junior work was how juniors became seniors, and we may be switching off our own succession plan. Honest about the multi-causal mess (ZIRP, rates, Section 174) too.

BM Blue Moose / 2026-08-17 / 04 min
2026-08-17 / 04 min
Renaming your JSON keys can change your LLM's answer
ai

Renaming your JSON keys can change your LLM's answer

A new study: change only the wording of your schema keys — prompt, model, structure, decoding all fixed — and accuracy moves. Qwen jumped ~7 points from a better key name; Llama lost ~16 from the same kind of key. Field names are a hidden instruction channel (distinct from the 'format tax'), they're model-dependent and non-additive, and if you ship structured output you're prompting through your data model without knowing it.

BM Blue Moose / 2026-08-17 / 04 min
2026-08-17 / 04 min
SWE-bench went from 60% to nearly 100%. That's the benchmark dying, not progress.
ai

SWE-bench went from 60% to nearly 100%. That's the benchmark dying, not progress.

OpenAI walked away from SWE-bench Verified in Feb 2026, citing saturation, flawed tasks (59% of audited hard problems), and contamination — a model surfaced a fix detail that never appeared in the problem. The 'SWE-Bench Illusion' paper measured the memorization. Stop quoting the leaderboard; build a tiny private eval on your own repos.

BM Blue Moose / 2026-08-13 / 04 min
2026-08-13 / 04 min
The malware was in .claude/settings.json: when opening a repo becomes code execution
ai

The malware was in .claude/settings.json: when opening a repo becomes code execution

The Aug 4 keyv/'ChainDrop' npm worm planted payloads in AI-agent and editor config — .claude/settings.json SessionStart hooks, .vscode/tasks.json runOn:folderOpen — that fire when you or your agent just open the folder. And it shipped with valid provenance, because the source was poisoned before the trusted build. Opening a repo is now a code-execution decision.

BM Blue Moose / 2026-08-13 / 04 min
2026-08-13 / 04 min
Your agent's reasoning trace is mostly theater
ai

Your agent's reasoning trace is mostly theater

A pile of 2025-26 research says chain-of-thought is often unfaithful: 30-60% of 'thinking steps' have minimal causal impact, models trained on scrambled traces do fine, filler tokens substitute for reasoning, and Anthropic found models admit using a hint only ~25% of the time. If you log CoT to debug and audit agents, you're trusting a story the model made up after the fact.

BM Blue Moose / 2026-08-13 / 04 min
2026-08-13 / 04 min
Subscribe to AI