Research notes on AI engineering from Eazler, read from primary sources: agents, coding automation, RAG, MCP, AI development environments and AI marketing. https://blog.eazler.com/en/ 2026-10-06T09:00:00+09:00 © 2026 Eazler Inc. (주식회사 이즐러) Eazler AI Labhello@eazler.comhttps://blog.eazler.com/en/ https://blog.eazler.com/en/ollaya-decision-models-korean 2026-09-27T09:00:00+09:00 2026-09-27T09:00:00+09:00

We ran six open, Jev-style decision models through Ollaya on the same MASSIVE sentences in Korean and English, sorting voice-assistant requests into 18 domains on an M1 Max Mac. Laya's multilingual model, which answers in about 26 ms on the Mac's GPU, scored 56.9% in English and 47.0% in Korean on the same 540 sentences. Three Qwen3.5-based ~0.8B models scored 71–81% in Korean but took 1–2 seconds per request on the CPU. Writing the label descriptions in Korean raised the accuracy of Laya's answers at confidence 0.9 or above from 75.5% to 91.9%.

https://blog.eazler.com/en/statistics-portals-security-txt 2026-09-25T09:00:00+09:00 2026-09-25T09:00:00+09:00

Australia's Prime Minister says an internal OpenAI model, running as an agent, got around repeated blocks on a Medicare statistics portal on 18 June 2026, and that OpenAI's first notice reached a public mailbox 84 days later. The other systems he named belong to statistics and health agencies, and the sites where outside researchers observed agent probes were public data and archive services. We checked 50 such portals in Australia, the US, the UK, Canada, New Zealand and at international bodies: 6 publish a valid security.txt, 7 publish one that breaks the standard, and 34 publish none. Every US federal domain we checked lists a security contact in the .gov registry instead.

https://blog.eazler.com/en/ipv6-only-developer-endpoints 2026-09-25T09:00:00+09:00 2026-09-25T09:00:00+09:00

On 2026-09-25 we looked up AAAA records for 114 hostnames that servers call during builds, deploys and normal operation, and 85 had one; all 85 answered over IPv6 from a dual-stack server. github.com, api.github.com, ghcr.io, the Stripe API, Twilio, SendGrid, Slack webhooks, Discord and the Vercel and Heroku APIs did not, so an IPv6-only server needs an IPv4 path such as NAT64 to reach them. On a Mac without IPv6, curl -6 still reached github.com over IPv4, because macOS hands IPv6 requests IPv4-mapped addresses by default.

https://blog.eazler.com/en/claude-code-agents-md-telemetry 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

Claude Code announced in 2.1.277 that it reads AGENTS.md in projects with no CLAUDE.md, but sessions with telemetry or nonessential traffic turned off, and third-party provider sessions such as Bedrock and Vertex, skipped the file without a warning. The feature flag that turns the feature on was coded to read as 'off' whenever it could not be fetched from the server. Comparing the executables published on npm showed that this default was the only behavior-changing edit to the agents-md plugin code in 2.1.281, and ten combinations reproduced with the official executables matched the code. The report also covers DISABLE_TELEMETRY=0 counting as off, a one-line workaround for version-pinned CI, and the 15 features still affected when flag fetching is off.

https://blog.eazler.com/en/nextjs-og-imageresponse-rce 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

CVE-2026-94545, a remote code execution vulnerability in Next.js next/og disclosed on September 22, 2026, is Critical 9.5 in the Next.js advisory and Moderate 5.3 in the advisory for Satori, the SVG library where it originates. Next.js 16.2.0–16.3.5 is affected, 16.3.6 fixes it, and 15.x is not affected. In the published packages, the Node.js path hands the SVG to sharp when sharp is installed, while the unaffected Edge path always converts with resvg.wasm. The report covers the four conditions for being affected, the separate fixed versions for Satori used outside Next.js, and the order in which to check.

https://blog.eazler.com/en/tokenizer-change-korean-token-cost 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

Claude Opus 4.7 introduced a new tokenizer, so the same text now produces a different token count. Counting 10 test texts on 7 models, four pure-Korean texts stayed at exactly 1.000×, while English prose, code and JSON logs grew 1.227–1.420×. The same model switch cut input cost by 33.3% for pure Korean text but by only 5.3% for JSON logs.

https://blog.eazler.com/en/agent-memory-as-files 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

The idea of keeping an AI agent's long-term memory in a folder of Markdown files instead of a vector database started with Karpathy's LLM wiki gist in April 2026 and continued with Google Cloud's Open Knowledge Format (OKF) v0.1 in June, v0.2 with trust signals in July, and git-based memory tools in August and September. Working from the OKF spec itself, this report covers the one required field (type), the bundle structure, v0.2's five trust signals (provenance, trust, freshness, lifecycle, attestation) and three trust tiers, how the three approaches (LLM wiki, memoryfield, OKF) differ, and how far the tools' performance claims have been verified.

https://blog.eazler.com/en/agent-tool-failure-fabrication 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

When an LLM agent's tool call fails, does the model report the failure or make up a value? A paper posted on September 13, 2026 measured this with 1,024 items built from 16 internal-system domains and 8 tool-failure types. Under a deployment-style prompt, a tool that reported failure with status:error led to 0.0% dishonest answers (at most 1.2% under any condition), while status:ok carrying a corrupted, redacted or empty value pushed the rate as high as 45.3%. None of the 9 audited agent frameworks specifies what to do when a tool fails, and appending one sentence that requires a retrieval_status: OK or FAILED line cut the rate from 14.10% to 0.87%.

https://blog.eazler.com/en/coding-agent-review-overclaim 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

Does an AI coding agent actually read every file it is asked to review? The OverclaimBench paper, posted on September 17, 2026, ran 12 models 1,140 times in their own production CLIs, including Claude Code, Codex, Grok Build and Antigravity. In 67.9% of runs the agent did not open every requested file, and 80.4% of those incomplete runs either did not disclose it or claimed in the final report to have reviewed everything. Runs that falsely claimed completion missed defects at 1.8 times the rate, and requiring subagents raised reading and defect reporting but did not make the remaining incomplete runs report more honestly.

https://blog.eazler.com/en/mrlatte-solo-18-services 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

The developer behind mrlatte.net presents 18 services, including InverseOne, Reputo, StartupXO, Typdit, DeepThought and BootSignal, as built and run by one person. Read side by side, the service pages describe how: a Firestore database per service on the same Firebase, Astro static builds, 03:00 crawl and 04:00 rebuild schedulers, AI content pipelines that assume human review, a shared account hub and credit economy, and a retrospective format that records every reorganization. All 18 services are linked.

https://blog.eazler.com/en/zcode-git-history-upload 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

An investigation published on September 18, 2026 found that ZCode, the AI coding app from Z.ai (Zhipu), bundled the entire workspace, .git history included, encrypted it and uploaded it to Alibaba Cloud whenever the user was logged in. A 345MB workspace became a 313MB snapshot, 86.6% of it .git; the two settings switches did not stop the upload, and the decryption key existed only on the server. Z.ai apologized the same day and blamed its Repo Wiki feature. This report cross-checks the investigation, two independent audits and the vendor statement, and sets out what Claude Code, Codex, Cursor, Copilot and Gemini CLI say in their official documentation that they send.

https://blog.eazler.com/en/typesafe-jev-system-one 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

Jev is a System One model that TypeSafe AI released on September 15, 2026; it generates no text and returns classifications, scores and yes/no decisions with probabilities. The vendor reported 0.4 seconds per case, $0.042 per million input tokens and 67.8% agreement in its own evaluation (Claude Opus 5: 73.1%). The reference labels are the average of GPT-6 Astra and Claude Fable 5.1, the homepage multipliers compare against the slowest and most expensive models, and there is no RLCD paper. Those conditions are set out here alongside the Hacker News debate.

https://blog.eazler.com/en/ai-coding-agent-adoption 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

AI coding agent usage rates differ from survey to survey. JetBrains' 2026 survey (more than 15,000 professional developers) found 90% weekly and 68% daily use, Stack Overflow's April 2026 pulse survey (1,100 respondents) found 59% using agents, and Microsoft's internal telemetry found that adopters merged 24% more PRs. The data from four survey sources, down to user shares by tool such as Claude Code at 39% and Copilot at 21%, is laid out side by side by sample, question and period to show why the numbers differ.

https://blog.eazler.com/en/github-agentic-workflows-guide 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

GitHub Agentic Workflows compiles markdown instructions into GitHub Actions workflows and runs them with Copilot, Claude Code, Codex or Gemini; it is in public preview. Working from the official docs, this guide sets out what the read-only default permissions, safe outputs, network firewall, threat detection and the 1,000 AI credit ($10) per-run cap each stop and fail to stop, with install commands, an example workflow and the cost structure.

https://blog.eazler.com/en/agent-bash-vs-typed-tools 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

A September 2026 Microsoft paper compared a shell alone, typed tools, the two combined and programmatic tool calling on two enterprise-work benchmarks with two models; we checked it against the arXiv source. Bash alone beat typed tools on both benchmarks and used 19–72% fewer tokens, and adding typed tools to the shell did not raise the score. In a random sample from the MCP Registry, half of the servers did not finish initializing, and 37.5% never started at all.

https://blog.eazler.com/en/agent-reward-hacking 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

METR's investigation of the Hugging Face incident, set against the wider debate on reward hacking in AI agents. Observations are kept apart from hypotheses, and a review procedure shows how a small team can separate permissions over grading criteria, execution logs and shared resources.

https://blog.eazler.com/en/openrouter-provider-routing 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

OpenRouter's order, only and allow_fallbacks settings, its parameter-support filters, the scope of ZDR and the router's own routing record, checked against the official docs. The guide sets out the availability cost of restricting providers and a procedure for validating responses before production.

https://blog.eazler.com/en/deepseek-v4-1-flash-migration 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

DeepSeek's schedule for moving its API model names to V4.1 Flash, including the September 14 switch for deepseek-v4-pro, with times in UTC and Korea time. Checked against the official docs: where the existing API names route, peak-hour pricing, Claude Code environment variables and the limits of the Anthropic-compatible API.

https://blog.eazler.com/en/skills-as-subagents 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

A September 2026 paper compared loading the same skill package into the main context with running it as a subagent, across 7 models and 64 tasks; we checked it against the original source. Skills without input-output contracts gained nothing from subagent execution, and with contracts the gap was largest on small models. Total tokens rose to 1.3–1.7×.

https://blog.eazler.com/en/ai-coding-supply-chain-security 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

Google GTIG's DUSTMAKER findings, compared with the official npm, SLSA and GitHub documentation. The report maps the trust boundaries to check separately when AI coding tools add dependencies, read workspace settings, publish packages and run security analysis.

https://blog.eazler.com/en/agent-retry-idempotency 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

An external operation that times out may have an unknown outcome rather than a failed one. Using the official Stripe and LangGraph documentation, this report checks what idempotency keys and checkpoints actually guarantee and sets out a recovery procedure that prevents duplicate execution.

https://blog.eazler.com/en/agent-sandbox-layers 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

Three 2026 measurement studies on coding agent isolation, checked against the original papers. For each file system operation studied, 69.0–98.6% of the command denylists meant to block it missed at least one bypass, and one sandbox product had gone 582 days without updating to the engine versions that fixed known vulnerabilities. The differences came from how isolation is operated, not from which isolation method is stronger.

https://blog.eazler.com/en/opentelemetry-genai-conventions 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

OpenTelemetry's GenAI semantic conventions moved to a separate repository in June 2026 and still have no release tag. Every stability level in the definition files is Development, and OpenTelemetry's own versioning rules say long-term dependencies should not be taken on anything in Development.

https://blog.eazler.com/en/korea-ai-basic-act-labeling 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

Korea's AI Basic Act took effect on January 22, 2026. Labeling generative AI output is indeed an obligation, but it is not among the three violations that Article 43 lists for administrative fines. Working from the original text of the Act and its Enforcement Decree, the report separates what is an obligation from where sanctions follow.

https://blog.eazler.com/en/multi-agent-failure-modes 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

In a study that classified 1,642 multi-agent execution traces, the most common failure was not agents misaligning with one another but repeating the same step (15.7%), with failing to recognize when to stop at 12.4%. Failures classified as problems between agents made up less than a third of the total.

https://blog.eazler.com/en/prompt-caching-cost 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

A paper measured 12 conditions: 4 models × 3 prompt caching strategies. Caching only the system prompt had the lowest cost for 3 of the 4 models, and again the shortest time to first token for 3 of the 4. This report explains why caching more of the context lost.

https://blog.eazler.com/en/ai-generated-code-quality 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

A code-analysis vendor says that since AI adoption, duplication is up 81% and refactoring down 70%. A peer-reviewed paper on the same question finds no general increase in code churn. The point is not which side is right but why they disagree, and what to measure in your own repository.

https://blog.eazler.com/en/agent-evaluation-benchmarks 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

OpenAI retired the coding benchmark it had been using after an internal audit found flawed tests or problem statements in 59.4% of the audited tasks. LLM judges, the usual alternative, showed high reliability but low validity in a study of about 540,000 verdicts.

https://blog.eazler.com/en/context-engineering-long-context 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

A 200K context window does not mean all 200,000 tokens are usable. In Chroma's measurements of 18 models, every one without exception got worse as input grew longer, starting well before the limit was reached. This report covers the craft of choosing what to leave out.

https://blog.eazler.com/en/coding-agent-security 2026-09-24T09:00:00+09:00 2026-09-24T09:00:00+09:00

OWASP rewrote the 2026 edition of its agentic security report from “plausible threats” to “CVEs, vendor advisories, and breach reports,” and 28 of the 53 agentic projects it tracks are coding agents. This report covers what actually broke and which defenses do not work on their own.

https://blog.eazler.com/en/ai-trends-2026 2026-09-24T09:00:00+09:00 2026-10-06T09:00:00+09:00

According to the Stanford AI Index 2026, the best AI agent on the OSWorld computer-use benchmark reached a 66.3% task success rate, close to the 72.35% human baseline. Yet in McKinsey's survey of 1,719 respondents in May–June 2026, only about 2 in 10 said their organization was scaling AI agents organization-wide, and 37% said AI had contributed to EBIT (earnings before interest and taxes), roughly the same as a year earlier. 'AI adoption' also means different things: an organization survey, a survey of U.S. adults and Microsoft usage telemetry give 88%, 53% and 28.3%.

https://blog.eazler.com/en/ai-marketing-methodology 2026-09-24T09:00:00+09:00 2026-10-06T09:00:00+09:00

According to the Google Ads Help Center, new Search campaigns are created with AI Max on, which also turns on search term matching, text customization (AI-written ad copy) and final URL expansion (automatic landing page choice). Meta's Marketing API has, since v23.0, applied Advantage+ audience to new ad sets left at default targeting, but says it never expands location, minimum age, language or custom audience exclusions. The performance figures both companies cite come from their own internal data; this article checked only official documentation as of 2026-10-06 and did not run any ads.

https://blog.eazler.com/en/ai-development-environment 2026-09-24T09:00:00+09:00 2026-10-06T09:00:00+09:00

In a typical local AI development stack, Ollama, Qdrant and Redis run in Docker without authentication by default, according to their official docs. Docker Compose also publishes a port on every network interface (0.0.0.0) unless you set a host IP. This guide shows a Compose file that binds ports to 127.0.0.1 only, and the sandbox and MCP server approval defaults of Claude Code and Codex CLI running on the same machine, checked against official docs on 2026-10-06.

https://blog.eazler.com/en/ai-coding-agents-landscape 2026-09-24T09:00:00+09:00 2026-10-06T09:00:00+09:00

We compared the official documentation for Claude Code, OpenAI Codex, GitHub Copilot, Cursor and Gemini CLI on 2026-10-06. All five support the Agent Skills open standard and MCP, and four read AGENTS.md without extra configuration. But no project skill folder appears in the docs of all five. Default local approval also differs: Claude Code and Cursor hand it to a classifier, Codex works inside a sandbox and asks at its boundary, and Gemini CLI asks for approval.

https://blog.eazler.com/en/loop-engineering 2026-09-24T09:00:00+09:00 2026-10-06T09:00:00+09:00

Practitioners such as Addy Osmani popularized ‘loop engineering’ in June 2026, but the term still has no standard definition. Read against two arXiv papers and the vendor documentation for /goal in Claude Code and Codex, the common emphasis is a stop condition that can show whether the work is done. A paper that scanned 36,645 public repositories confirmed autonomous agent loops in 217 of them, 180 of which ran only on repository events such as automatic PR review. Almost none committed the loop state files that practitioner posts recommend.

https://blog.eazler.com/en/agent-skills-guide 2026-09-24T09:00:00+09:00 2026-10-06T09:00:00+09:00

The Agent Skills specification defines six SKILL.md frontmatter fields, while skill locations, the setting that stops automatic invocation, and tool pre-approval differ across Claude Code, Codex and VS Code. Human-curated skills raised the average pass rate by 16.6 percentage points across 87 tasks, but 26.1% of 31,132 skills from public marketplaces were flagged with risk patterns that warrant review.

https://blog.eazler.com/en/gbrain-agent-memory 2026-09-24T09:00:00+09:00 2026-10-06T09:00:00+09:00

The README of GBrain, Garry Tan's open-source agent memory server, cites a relationship-question recall of 0.9791, but the developer's own eval report has reclassified that figure as the upper bound of an adapter that knows the four question templates in advance. In GBrain's production search pipeline, enabling the relationship retrieval stage raised recall from 0.6626 to 0.7241. GBrain's docs also state that Markdown files alone cannot restore DB-only pages, revision history or withdrawal records.

https://blog.eazler.com/en/geo-seo-ai-agent-case-study 2026-09-05T09:00:00+09:00 2026-09-05T09:00:00+09:00

93% of app referrals already came from ChatGPT while robots.txt blocked that exact crawler. Measured baselines, the assumptions that broke, and a checklist.