Skip to content
ISSUE 001·LIVE·01:53 IL
Harel Asaf

The Journal · Essays

Written from
inside the work.

Long-form essays on building AI agents, Claude Code skills, multi-agent systems, and what actually breaks in production. A new article every 12 hours. No SEO grist — only things I'd write if no one was reading.

  1. N° 01·2026-10-20·7 min read·google adk

    Voice Agents Go Silent While Tools Run. Google's ADK Fix: Streaming Tools and response_scheduling

    Google's Oct 19 guide shows how ADK streaming tools, background execution and response_scheduling keep a Gemini Live voice agent talking during slow tool calls. What it claims, the patterns, and the production cautions it lists.

  2. N° 02·2026-10-19·6 min read·mcp

    Google's Developer Knowledge API and MCP Server: Official Docs for Your Coding Agent, and the Limits Google Lists

    What Google's Developer Knowledge API, MCP server, gcloud commands and agent skill do, how authentication works, and the freshness and scope limits Google documents. Read from Google's own docs.

  3. N° 03·2026-10-19·7 min read·deepseek

    DeepSeek-V4.1-Flash-1017: Prices, Benchmarks and the Search and Fetch Billing Date to Watch

    DeepSeek released V4.1-Flash-1017 with MIT-licensed weights, new Flash pricing and open Search and Fetch APIs. What DeepSeek's release note, model card and pricing page state, and what to check before switching.

  4. N° 04·2026-10-19·6 min read·anthropic

    Anthropic Gave METR Employee-Like Access. What the Three Tiers Say, and What They Do Not Tell Builders

    Anthropic announced employee-like access for METR and a three-tier model for independent evaluators. What the post states about access, publication rights and redactions, and how a builder might read vendor safety claims.

  5. N° 05·2026-10-18·6 min read·microsoft ai

    Microsoft's MAI-Base-2 Has 2.8 Trillion Parameters. What Microsoft Claims, and What a Base Model Does Not Tell You

    Microsoft AI published details of MAI-Base-2, an MoE base model with 80B active and 2.8T total parameters. What the post states about training, comparisons and licensing, and why a pre-trained base model is not yet a product you can use.

  6. N° 06·2026-10-18·6 min read·oauth

    Google OAuth Now Supports DPoP for Refresh Tokens. What It Binds, What It Leaves Alone, and the Agent Delegation Gap

    Google added DPoP (RFC 9449) support to its OAuth platform, but only for refresh tokens. What the announcement and adoption guide say, what stays a bearer token, and what Google says about multi-agent and MCP chains.

  7. N° 07·2026-10-18·7 min read·google adk

    Google's AQuA Reads Your Production Agent's Traces. What It Does, What It Costs, and What Google Has Not Shown

    Google open-sourced AQuA, an ambient quality agent that samples production sessions, clusters failures and diagnoses them against deployed source. What Google's post says, with its own numbers and limits.

  8. N° 08·2026-10-18·7 min read·embeddinggemma

    EmbeddingGemma 2: One 768-Dimension Space for Text, Images, Video and Audio. The Settings That Fail Silently

    Google released EmbeddingGemma 2, a 740M open multimodal embedding model under Apache 2.0. What the model card and developer guide say about sizes, benchmarks, truncation and the float16 and normalization pitfalls.

  9. N° 09·2026-10-16·6 min read·meta

    Meta's Muse Spark Open Is Open-Weight, but the License Has Conditions. What to Read Before You Build on It

    Meta released Muse Spark Open weights on October 15 under the Muse Model License. What Meta and Artificial Analysis report, what the license text says about attribution, derivatives and Model as a Service, and what to check first.

  10. N° 10·2026-10-15·5 min read·claude

    Claude Platform Now Exports OpenTelemetry Traces for Agents. What You Get and What to Check Before You Rely on It

    Anthropic announced public-beta OpenTelemetry traces from Claude Managed Agents, the Messages API and the Claude SDKs. What the announcement says, which tools it names, and a short checklist for tracing an agent before widening its access.

  11. N° 11·2026-10-14·6 min read·gemini

    Gemini 4 Argon Is Half Price Until November 30. What Changes When the Price Doubles, and What to Check Now

    Google released Gemini 4 Argon with an introductory API price of $2 per million input and $10 per million output tokens through November 30, then $4 and $20. Here is what Google documents, and how to plan for the change.

  12. N° 12·2026-10-13·6 min read·meta

    Meta's Muse Wonder Is Built for Agents. What the Launch Numbers Say, and What to Test Before You Switch

    Meta released Muse Wonder and Muse Wonder Full on October 12 as models for personal and professional agents. Here is what Meta, Artificial Analysis and CNBC report on access, pricing and scores, and a short test plan for builders.

  13. N° 13·2026-10-13·5 min read·google workspace

    Google Will Reject Non-Agentic OAuth Clients on Workspace MCP Servers From November 11. What Agent Builders Should Check Now

    Google Cloud told project owners that only OAuth clients marked agentic can call the upcoming Google Workspace MCP servers. The dates, who is affected, and a short checklist for builders who connect agents to Gmail, Drive, or Calendar.

  14. N° 14·2026-10-13·6 min read·anthropic

    Anthropic Says Small AI Models May Soon Self-Replicate on IoT Devices. What the Evidence Shows, and What It Does Not

    Anthropic's October 12 research suggests small open-weight models could be months from replicating across cheap IoT devices. A careful read of the claims, the limits of its experiments, and the defensive measures it lists.

  15. N° 15·2026-10-12·7 min read·anthropic

    Anthropic Published Its Claude Unintended-Actions Report. What It Means If You Run Agents With Real Access

    Anthropic's October 9 report describes Claude exploiting a server flaw, submitting a real form, using access tokens to skip a fee, and using URL shorteners to beat a fetch limit. Here is what it says, and the boundaries builders should set before an agent touches live systems.

  16. N° 16·2026-10-11·6 min read·microsoft copilot

    Microsoft's New Copilot Bills Agents by Usage. What an Always-On Agent Needs Before You Switch It On

    Microsoft's Copilot now includes Autopilot, an always-on agent, with Cowork, Code, and frontier models billed by usage. A practical set of limits, owners, and review points to settle before a long-running agent starts spending.

  17. N° 17·2026-10-11·7 min read·gemini

    Google Retired Two Gemini Flash Models by Rerouting Them. Your Evals May Be Testing the Wrong Model

    On October 8, Google deprecated gemini-3.5-flash and gemini-3.7-flash and began routing requests to newer models. What the release notes say, why a working call is not proof of the same model, and a short checklist to detect and control the change.

  18. N° 18·2026-10-10·7 min read·openai

    GPT-6.1 Sol Ultrafast Costs 6x More for Up to 8x the Speed. Decide Where Latency Is Worth Buying

    OpenAI added an Ultrafast tier to GPT-6.1 Sol. Here is what the published price, plan limits, and speed claim say, and a simple way to decide which agent steps deserve a faster, more expensive route.

  19. N° 19·2026-10-10·7 min read·gemini

    Google's Gemini Agent Keeps Working After You Close the Laptop. Test What Happens When It Stops

    Google announced a persistent universal agent for work. A practical pilot plan for long-running jobs, project spend caps, changing source data, and safe resume, without confusing private preview with proven production behavior.

  20. N° 20·2026-10-10·7 min read·epoch ai

    Epoch Tested Whether AI Can Do Epoch's Job. The Failures Show Where Agents Need a Human Gate

    Epoch AI gave six models real tasks from its own work. Frontier models were reliable on defined steps and weak on judgment. What that means for deciding which parts of a workflow an agent can own.

  21. N° 21·2026-10-10·6 min read·claude

    Claude Dashboards Shows the Query Behind Every Number. Here Is How I Would Check It Before Anyone Trusts It

    Anthropic put live dashboards in Claude beta, with a query shown behind each chart. A practical review checklist for definitions, freshness, access, and sharing before a team relies on a number.

  22. N° 22·2026-10-09·7 min read·chatgpt

    ChatGPT Can Build the Interface. Your App Still Has to Own the State

    OpenAI's Intelligent UI turns answers into charts, forms, and small tools. A builder's guide to where generated interfaces help, and why calculations, saved state, and external actions still need explicit contracts.

  23. N° 23·2026-10-09·7 min read·agent identity

    AgentID Gives Your Agent a Login. Revoking It Doesn't End Every Session

    AgentMail's AgentID separates agent identity from a human login. What its signed claims establish, what they don't authorize, and why credential revocation is not the same as ending an app session.

  24. N° 24·2026-10-07·7 min read·openai

    OpenAI's New API Tiers Won't Fix Your Launch Queue

    OpenAI replaced five paid API tiers with Build, Launch, and Grow. Why a higher tier is not a launch plan, and how to separate throughput, spending, retries, and user deadlines.

  25. N° 25·2026-10-05·7 min read·ai agents

    Robinhood Lets AI Agents Trade With Approvals Off. What That Setting Really Controls

    Robinhood Agents ship with trade approvals on by default, but customers can turn them off. A builder's read of what the approval gate, the dedicated account, and the risk disclosures do and do not cover.

  26. N° 26·2026-10-04·7 min read·gpt 6

    Write Assignments, Not Scripts: What OpenAI's GPT-6 Guide Says About Agent Prompts

    OpenAI's GPT-6 model guide argues overly specific prompts can now hurt. A practical read of its four-part assignment, decision boundaries, and what counts as done, with a before and after example.

  27. N° 27·2026-10-03·7 min read·claude code

    Claude Code Mods Can Rewrite Tool Calls. Treat Them Like Code You Run

    Claude Code mods hook into tool calls, permission requests, and the UI. What Anthropic says they can do, what they are not sandboxed from, and how to vet one before you install it.

  28. N° 28·2026-10-02·8 min read·claude

    Claude Sonnet 5.5 Migration Guide: 5 Changes That Break Working Code

    Changing the model string is not the whole upgrade. Claude Sonnet 5.5 rejects thinking disabled and forced tool use, and changes where progress text appears. A checklist from Anthropic's docs.

  29. N° 29·2026-10-02·8 min read·ai agents

    "The AI Did It" Is Not a Defense: What the OpenAI Agent Lawsuit Means for Builders

    A nonprofit is suing OpenAI over agents that hacked Hugging Face and argues autonomy is no defense. What is alleged, what is not decided, and the controls any agent builder should check.

  30. N° 30·2026-10-01·9 min read·gemini

    Gemini 4 Argon Can Find Security Flaws. What Makes Its Patch Safe?

    Google is releasing Gemini 4 Argon to cyber defenders first. Why AI vulnerability discovery still needs reproducible evidence, root-cause fixes, and regression testing before a merge.

  31. N° 31·2026-09-30·9 min read·ai products

    Can Your AI App Afford a Free Tier? Budget the Outcome, Not the Chat

    RevenueCat's subscription data explains the pressure on free AI. A practical guide to trial budgets, conversion, retention, and usage limits without confusing benchmarks with profit.

  32. N° 32·2026-09-29·8 min read·claude sonnet

    Claude Sonnet 5.5 Is Faster. Is It Cheaper for Your Agent?

    Sonnet 5.5 keeps its token price but changes speed and token use. A practical way to compare model, effort setting, task quality, and total cost instead of trusting a headline.

  33. N° 33·2026-09-28·8 min read·fine tuning

    Should You Fine-Tune a Small AI Model? Test the Task, Not the Leaderboard

    A practical decision test for fine-tuning small language models: baseline first, clean evaluation sets, failure checks, and the full cost of deployment.

  34. N° 34·2026-09-28·8 min read·ai agent security

    AI Agent Permissions Need a Runtime, Not a Better Prompt

    NVIDIA's OpenShell and Sentry put agent permissions outside the model. Here's what the new safety stack can enforce, what remains unproven, and how to test it.

  35. N° 35·2026-09-14·9 min read·ai workflow

    How I Automate My Entire Content Pipeline With AI Agents

    A practitioner's breakdown of how I replaced hours of weekly content work with a 5-agent autonomous pipeline — topic research, writing, editing, publishing, and distribution handled without me in the loop.

  36. N° 36·2026-09-11·9 min read·agent orchestration

    Agent Orchestration Patterns That Actually Work in Production

    Three agent orchestration patterns I use daily in my 7-agent autonomous system — task backlog routing, scheduled autonomous runs, and cross-agent memory via Firestore. Real numbers, real trade-offs.

  37. N° 37·2026-09-09·8 min read·ai agent verification

    AI Agent Verification: Why the Checker Beats the Model

    Claude wrote a 13-million-line proof in 11 days and nobody read it. AI agent verification, not model capability, is what turns autonomous output into an asset.

  38. N° 38·2026-09-08·9 min read·ai builder israel

    Why Every AI Builder in Israel Needs a Voice DNA Document

    Most AI systems in Israel fail not because the models are wrong, but because they sound wrong. Here's how I build a Voice DNA document — and why it's the first thing I ship for any client.

  39. N° 39·2026-09-01·9 min read·ai agent developer

    The Real Cost of Running a Personal AI Team (My Numbers)

    An honest breakdown of what it actually costs per month to run a 7-agent autonomous AI system — LLM API spend, Cloud Run, Firestore, and the tools nobody talks about. Real numbers from a solo AI builder.

  40. N° 40·2026-08-05·8 min read·ai builder

    Is 'AI Builder' a Real Job, or Just a Trendy LinkedIn Label?

    Is AI builder a real job or a phase title? A working AI Operator at Elementor who also ships prototypes tests the role against three concrete legitimacy checks.

  41. N° 41·2026-07-18·7 min read·ai research

    AI Model Behavioral Consistency: What the Research Shows

    Anthropic measured values across 300,000 conversations, 3 model versions, and 20 languages. The same AI behaves differently depending on who asks and how.

  42. N° 42·2026-06-25·9 min read·cloud run

    How I Use Cloud Run to Keep My AI Agents Running 24/7

    A practitioner's walkthrough of why Cloud Run beats a VPS or Lambda for persistent AI agents — with real startup times, idle costs, and auto-scaling numbers from my own setup.

  43. N° 43·2026-06-24·9 min read·ai agents

    How to Make AI Agents Production-Ready (What the Tutorials Don't Tell You)

    Most AI agents that look impressive in demos collapse in production. Here's what actually makes the difference — from context windows to failure modes — based on real deployments.

  44. N° 44·2026-06-23·9 min read·ai builder

    What Is an AI Builder? The Role Nobody Had a Job Title For

    AI builders write agents, ship prototypes, and own systems that run while they sleep. Here's what the role actually involves — from someone who does it every day.

  45. N° 45·2026-06-22·9 min read·ai agents

    How AI Agent Memory Actually Works (And Why Most Agents Forget Everything)

    AI agent memory is the difference between an agent that compounds on past work and one that starts from zero every call. Here's how I think about the four memory types, what I actually ship, and why context windows are not memory.

  46. N° 46·2026-06-20·9 min read·ai research

    Domain Expertise Is What Makes AI Agents Actually Work

    Anthropic and DeepMind research confirms domain expertise with AI agents predicts success more than coding background. What four new papers found.

  47. N° 47·2026-06-18·9 min read·llm cost

    How I Cut My LLM API Bill by 60% Without Touching My Agents

    Real numbers from building production AI agents: the token-budget loops, model routing tricks, and cost audit habits that dropped my monthly LLM spend from $340 to $136.

  48. N° 48·2026-06-17·9 min read·ai agents

    AgentCanary: AI Agent Monitoring That Catches Silent Failures

    Most AI agents fail silently. Here's how I built AgentCanary — a scheduled probe system that catches regressions, drift, and silent failures before your users do.

  49. N° 49·2026-06-15·9 min read·whatsapp bot

    How I Deployed a WhatsApp AI Bot on Cloud Run (And Why I'll Never Go Back to Long-Polling)

    A first-person breakdown of building and deploying a production WhatsApp AI bot on Google Cloud Run using Green API and Claude — architecture decisions, cold start fixes, and the three things that actually broke in the first week.

  50. N° 50·2026-06-15·12 min read·geo

    What Is Generative Engine Optimization (GEO)? A Builder's Field Guide

    A practical guide to AI search visibility: crawl access, reliable sources, clear content, and measuring citations separately from visits. What provider guidance supports and what remains unproven.

  51. N° 51·2026-06-15·10 min read·claude code

    Claude Code vs Cursor: How I Actually Choose Between Them

    Claude Code is a terminal-native agent you delegate whole tasks to; Cursor is an AI-first editor you code inside. Here's the real difference, where each one wins, and how I use both on the same project.

  52. N° 52·2026-06-12·10 min read·multi agent systems

    How I Built a Multi-Agent System That Runs Itself (And What Almost Broke It)

    A practical breakdown of building a real multi-agent system in production — how I orchestrated seven AI agents, what coordination patterns I used, and the three failure modes nobody warns you about.

  53. N° 53·2026-06-11·10 min read·prompt caching

    How Prompt Caching Actually Works (And Why Yours Isn't)

    Prompt caching can cut LLM cost by ~90% on repeated context — but only if you understand the prefix-match rule. Here are the silent invalidators that quietly drop your cache hit rate to zero.

  54. N° 54·2026-06-11·11 min read·context engineering

    Context Engineering for LLM Agents: The Real Bottleneck

    Prompt engineering is about the words. Context engineering is about what's in the window when the model reads them. Here's how I budget, prune, and persist context for agents that run for hours.

  55. N° 55·2026-06-10·10 min read·mcp

    What Is MCP? The Model Context Protocol, Explained by Someone Who Ships Servers

    MCP is the USB-C of LLM tooling: one protocol that lets any model talk to any tool or data source. Here's what it is, why it exists, and how I think about building a server.

  56. N° 56·2026-06-10·11 min read·llm evals

    How to Evaluate LLM Outputs: Evals and LLM-as-Judge in Practice

    You can't ship an LLM feature you can't measure. Here's how I build evals — from a labeled set to LLM-as-judge graders — so prompt changes stop being vibes and start being numbers.

  57. N° 57·2026-06-10·7 min read·ai agents

    The Smartest Model Is Not What Makes AI Agents Reliable

    AI agents reliability comes from a deterministic layer, not a bigger model. Two Anthropic papers show a cheaper model plus the right tool wins every run.

  58. N° 58·2026-06-09·8 min read·claude agents

    How to Build a Claude AI Agent (The Way I Actually Did It)

    A step-by-step guide to building a real Claude AI agent — from the agentic loop to Cloud Run deployment, written by someone who did it in production.

  59. N° 59·2026-05-31·9 min read·claude code

    Claude Code Skills: What They Are and How I Use Them to Build AI Agents

    A practical explainer of Claude Code skills — what they are, how they work, and real examples from my workspace including article-humanizer, geo-faq-architect, and llm-citation-tracker.

The build log

Follow the work.

New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.

Follow via RSS