Insights on AI automation, Hyros attribution, and n8n workflows
n8n finally caches Claude prompts natively. Turn it on for every tool-calling AI Agent and leave it off for one-shot workflows, where it writes the cache every run and never reads it. The math and the fixes.
Read Full Article
Claude only fetches URLs that came from a user message, a search result or your own tool. System prompt, its own output and MCP results don't count, and Managed Agents got stricter on Oct 7. The fix for each case.
Read Article
$100 on Max 5x, $200 on Max 20x, every billing cycle. Not usable in Claude Code, no rollover. I priced a real agent run against it: 95 runs on Sonnet 5.5, 48 on Opus 5.5, 26 on Fable 5.1.
Read Article
On by default since n8n 2.43.0, it lets the Assistant drive your logged-in Chrome with 32 tools. Not the Browser Use node. What it can touch, why the extension won't connect behind a proxy, and the env var that turns it off.
Read Article
10x cheaper per token, 75% cheaper per task. I ran four real tasks on both: 35-39% more tokens for the same text, 8x the output on summaries, and a 100K cliff an 89K prompt now crosses. Effort low claws it back to 87%.
Read Article
Empty body or 'error code: 1010'? That tells you whether ComfyUI or Cloudflare said no. Nine live requests against my own box: the nginx Host bug, the Sec-Fetch-Site block that breaks web apps, and why the origin check isn't auth.
Read Article
claude plugin eval scores your skill with the plugin and without it. Four real runs on an n8n review skill: a vague description fired 2 of 5 times, 30 runs cost $1.03, and the default Haiku judge overstated the delta.
Read Article
Claude's browser toolset reads the page through an accessibility tree; the computer toolset only sees screenshots. Same tasks, same Chromium, Sonnet 5.5: 2 turns vs 6 on a long table, but the bigger browser definition lost on a simple form.
Read Article
Retell's browser SDK 2.x and /v2/create-web-call are deprecated on Oct 18. v3 hands out a gateway token SDK 2.x can't use, so the order matters. Tested v2 vs v3 side by side, plus how to rebuild the transcript and talking indicator v3 calls no longer send.
Read Article
The Code node timeout isn't slow code: the broker never matched your task to a runner. Reproduced in Docker on n8n 2.38.3 with the broker on 127.0.0.1, a wrong auth token and the working config, plus the six causes and the runner log line for each.
Read Article
Mods shipped in Claude Code 2.1.287. Settings hooks spawn a script per event; mods run TypeScript in-process and can draw panes, rewrite prompts and approve calls. Same .env guard built both ways, tested headless, with timings and the ordering trap.
Read Article
The Files API went GA. Removing the beta header swaps has_more for next_page, turns after_id into a 400 and adds expires_at plus 1-hour-to-90-day file expiration. The live tests, the silent pagination bug and the migration order.
Read Article
Sonnet 4.5 stops answering on the Claude API on 30 November. Sonnet 5.5 is cheaper per token, but prefill, temperature, budget_tokens, disabled thinking and forced tool_choice all return 400, and the default now thinks first. The exact errors, my cost test and the checklist.
Read Article
Retell's legacy LiveKit SIP host was retired on 30 September. Only the hostname changes, but it lives in your carrier trunk, not in Retell, and on retirement morning the old host still answered OPTIONS pings over UDP, TCP and TLS. Where to look, what the new host looks like, and how to prove the cutover with a real call.
Read Article
Cache diagnostics went GA on 23 September: send the previous response id and the API names the first thing that changed. I broke a 21,000-token cache five ways on purpose; a timestamp, reordered tools and one trailing space each cost every cached token, and a system-prompt breakpoint cut the history miss to 135 tokens.
Read Article
The caller's clock is the real limit, not n8n: a 110-second wait still answered on 2.38.3. What actually breaks is a branch that skips the response node (empty 200), an error before it (generic 500), and a slow node in front of it. The ack-first pattern that fixes all three.
Read Article
An empty {} from /history/{prompt_id} is four different answers: still queued, rejected at submit, wiped by a restart (history lives in memory), or evicted past 10,000 entries. The /queue check that tells them apart, and why I verify the file instead of the prompt_id.
Read Article
Cold drops the caller on a stranger. Warm waits for a human and whispers the context first. Agentic warm sends a second AI to ask if they can take the call — my default for anyone answering on a cell, because a voicemail greeting also counts as someone speaking. Plus the caller ID setting that makes transfers fail outright.
Read Article
Redaction is a view filter, not a delete. It hides payloads when someone opens an execution, but the rows stay unencrypted in Postgres, in your backups, and in any console.log a Code node printed. What it covers, what leaks, and the free save settings that actually keep customer data off disk.
Read Article
Two bills: runner minutes at $0.006/min, and tokens. The minutes are rounding error. The tokens cost more than on your laptop because a fresh runner never hits the prompt cache — so every run pays the same cold-start floor whether the PR rewrote 400 lines or fixed a typo. Which makes `synchronize` the multiplier nobody budgets, and four lines of YAML the fix.
Read Article
Cheaper on every line — $4/$20 per MTok, cache reads down 60% to $0.20 — same 1M context, same 128K output. Then four request shapes start returning 400, and the default effort level quietly drops from high to medium. The migration in order, the cost math that isn’t 40%, and why progress updates vanished from your agent UI.
Read Article
v2.1.278 moved auto mode’s safety classifier to the server and stopped charging for it — except in the sessions most likely to run it all day. Why an LLM gateway drops you back onto the billed path, the /status row that tells you in ten seconds, what actually skips the classifier, and why subagent fan-out is the multiplier nobody budgets for.
Read Article
Claude Code finally reads AGENTS.md — conditionally. The per-path fallback rule, the gitignored CLAUDE.local.md that switches it off for exactly one person on your team, the four session types where the feature isn't there at all, why it never appears in /context, and why the @AGENTS.md import beats a symlink every time.
Read Article
Five places to put instructions and one bad default. Where an output style belongs versus CLAUDE.md, the keep-coding-instructions field that quietly strips Claude Code's engineering behavior from every custom style, why switching styles keeps your prompt cache while editing CLAUDE.md does nothing until you restart, and why your subagents never see any of it.
Read Article
One conversation coordinates, and every task becomes a full Claude Code cloud session with its own context window, branch and pull request. What a thread is that a subagent isn't, why a new project runs everything on Opus at high effort, the 200-threads-a-day ceiling, and the three settings to change before you send it real work.
Read Article
$0.08/min buys agent hosting on ElevenLabs — the LLM and the phone line are billed separately on top. Retell's $0.07 is an itemized stack you assemble. The worked math on 3,000 minutes at 25 concurrent calls, why concurrency limits (not per-minute rates) pick the platform, and why swapping the model moves the bill harder than swapping the vendor.
Read Article
Claude picks a skill from its description alone — the SKILL.md body loads only after it's chosen. Why descriptions fail, what when_to_use is actually for, and how claude plugin eval's tool_used: Skill grader plus a no-plugin baseline turns "is my skill firing?" into a number you can gate CI on.
Read Article
Barge-in fires on audio energy, not on meaning — so a cough ends the agent's turn exactly like a real objection. What Retell's interruption_sensitivity and responsiveness each control, why Vapi's numWords is the knob that actually fixes it, and the tuning order that stops you shipping an agent that is both slow and rude.
Read Article
Terraform's plan-and-apply loop, for Claude Managed Agents. Write each agent, skill, environment and deployment as a file, commit the lockfile, and review prompt changes like code — plus the rename that quietly doubles your agent count.
Read Article
Claude has two memory systems now and they are not the same product. Memory stores are managed, versioned files Anthropic mounts at /mnt/memory; the memory tool runs on your own disk. The split, the setup, the read_write footgun, and the limits that fail quietly.
Read Article
Swapping an agent's tools used to nuke the prompt cache for the entire conversation. tool_addition and tool_removal fix that — here are the block shapes, the defer_loading twist, and the placement rules that return a 400.
Read Article
They fix opposite halves of the same token bill. How defer_loading and allowed_callers differ, when each is the wrong answer, and the MCP catch that quietly decides it for most stacks.
Read Article
A second opinion without a handoff. How the advisor server tool works, the billing field that hides advisor spend from your cost dashboard, and the pairing rules that 400 your request.
Read Article
A grader in a separate context window is the whole trick. The Outcomes API, why max_iterations is a budget and not a quality dial, and the check my own pipeline skipped 20 times while reporting success.
Read Article
Guard prompts fail. Tool scoping works. The four-layer defense I applied across 182 production workflows — and the nine exfiltration paths I'd built myself.
Read Article
npm installs die, AI Agent v1 is gone, three legacy nodes are removed. I audited 182 production workflows — only one actually needed a rewrite.
Read Article
Same $10/$50 per token, cache reads cut 75%. Your bill falls by 0.75 times your cache-read share — and three breaking changes can take it right back.
Read Article
Decide by who owns the agent loop. And since AI SDK 7, HarnessAgent runs Claude Code behind the AI SDK — so it stopped being an either/or choice.
Read Article
temperature is gone, httpx became httpx2, and Text Completions was deleted outright. The full v1.0 migration, with the fix for each change.
Read Article
A failed tool inside an AI Agent usually isn't a failed execution — so the Error Trigger stays silent. The exact code path, and how to make failures loud again.
Read Article
Checkpoints miss shell edits, background subagents, and every symlinked path. The coverage map, and the one habit that covers all four blind spots.
Read Article
One skill folder that loads unchanged in Claude Code, Codex, Cursor and Gemini CLI. The directories, the symlink layout, and the four things that break portability.
Read Article
Every tool call becomes a persisted, retryable step, so a crash or redeploy resumes instead of restarting. How DurableAgent and hooks work, and where n8n still wins.
Read Article
Subagents delegate and report back. Agent Teams shares a task list and git worktrees between sessions. They solve different problems — my decision framework, and the merge trap teams hide.
Read Article
The 5.5% fee is the smallest number in this decision. Break-even math, failover that only matters when it fires, and the cache mistake that costs 10x the fee.
Read Article
November 30, 2026 is the hard stop, and Evals goes read-only a month earlier. Export to Agents SDK code now — then sort each workflow between the SDK, Workspace Agents and n8n you actually own.
Read Article
Five levels, one dial — and it moves tool calls as well as thinking. Where to start per model, why Opus 5 reversed the advice from Opus 4.8, and the trap that quietly kills your prompt cache.
Read Article
Anthropic replaced vector search with grep in Claude Code. Most agents ship with a vector database they don’t need — here’s where agentic search wins, where RAG still wins, and the test that settles it.
Read Article
The registry hosts metadata, not code — so it won't take your server unless the published package proves it owns the name. Add mcpName before your first npm publish, not after.
Read Article
The “25x more tokens” stat everyone quotes ignores cache reads, which makes it about 4-5x, or cents per run. The real decision is durable execution versus an inherited harness.
Read Article
No surcharge isn’t the same as free. The per-turn cost math, the tokenizer change that quietly broke everyone’s estimates, and why a loop allowed to grow to 1M gets worse at its job.
Read Article
Data Tables are for rows that only serve your workflows — Postgres is for everything a human will eventually query. The 200 MiB instance-wide cap, the Code node gap, and the export that leaves your table behind.
Read Article
Metadata Filter only exists in Get Many mode — so an agent using a native vector tool searches your whole collection every time. Why pgvector wins by default, and the three ways to get scoped retrieval anyway.
Read Article
Streaming is two switches, not one — and n8n never tells you when one is off. The six reasons your agent still returns one blob at the end, including the buffering proxy that makes it work on localhost and fail on your domain.
Read Article
The approval link is a bearer token, the gate belongs to the tool rather than the agent, and a waiting execution never turns red. Five ways n8n human-in-the-loop approvals fail in production — and the gate I run instead.
Read Article
Webhook retries are at-least-once by design, and the Remove Duplicates node is a filter rather than a lock. Here is the atomic idempotency gate that holds under queue mode — plus the two-phase trick that stops it turning duplicates into silent data loss.
Read Article
Workflow limits are gone, so the old reason to self-host died with them. What you actually buy now is executions and concurrency — with real numbers pulled off my own production instance and the one plan cliff that decides everything.
Read Article
Access asks every caller to prove it is a person, and a webhook has no answer, so the delivery dies at the edge. Service tokens with a Service Auth policy, a Bypass scoped to the webhook path, and the split-hostname setup that ends the problem for good.
Read Article
Your container writes, the sharing layer reorders, and one morning the storage engine can't read its own file. Why bind mounts across the Docker Desktop VM boundary break Postgres and Mongo, the four-step migration to named volumes, and how to put Docker on an external SSD without a blank disk image after reboot.
Read Article
Your MCP tool works locally and dies behind the proxy. The Tasks extension answers tools/call with a claim ticket instead of a result — taskId, pollIntervalMs, five states — plus what breaks if you built against the 2025-11-25 experimental version.
Read Article
An agent buys through your API, not a browser, so no pixel ever fires and the revenue lands in your reports as nothing. Two separate failures, only one fixable — plus the server-side pipeline that recovers it and the part no vendor can sell you back.
Read Article
One exposes your entire n8n instance to an AI client, the other exposes a single workflow. Since April 29 2026 the built-in server can build workflows, not just run them — plus the silent version gates and the queue-mode routing rule that keeps SSE from dropping.
Read Article
output_config.format constrains the answer, strict:true constrains the tool call — they're not alternatives. Plus the unsupported keywords that make your Pydantic model bounce, and why changing the schema nukes your prompt cache.
Read Article
Server-side summarization for long Claude conversations — the exact config, the one line that silently breaks it, the billing trap that under-reports a compacting turn by 8x, and when context editing is the cheaper choice.
Read Article
SEP-2577 deprecated Sampling, Roots and Logging in spec 2026-07-28. Elicitation survived — here's why, the delivery change that breaks blocking tool handlers, and the phishing attack sitting inside URL mode.
Read Article
It's three limits, not one — and the fix almost nobody mentions is prompt caching, because cached input tokens don't count toward your ITPM ceiling. Plus the header you're throwing away that would let you throttle before the 429 ever fires.
Read Article
Both let a manager agent delegate to specialists. Only one gives you an execution you can retry, reuse and test on its own. That's the whole decision — plus the sub-node trap that makes both silently process only your first item.
Read Article
Writes are locked to your working directory. Reads are not — a sandboxed command can still open ~/.ssh and ~/.aws/credentials. Here's the real boundary, the settings.json that closes the gap, and the five options that quietly undo it.
Read Article
An MCP tool description is prompt text your agent reads as instructions and you never see — which is why no system prompt fixes it. The four structural controls that do: per-tool allowlists, pinned versions with description hashes diffed in CI, credentials the model can't read, and a container with --network none.
Read Article
A marketplace is a git repo with one JSON file in it — no server, and your git host is the access-control layer. The full setup, plus the two traps: private HTTPS auto-updates that fail silently, and the optional version field that ships every commit to your whole team.
Read Article
n8n gates Git source control behind the Business plan at roughly €800/month. The public REST API ships on every plan and covers most of it — export, normalize so the diffs are readable, commit on a cron. Plus the two things no plan can save you: credentials and the restore path.
Read Article
n8n tells you what a node did. It won't tell you why the agent picked that tool, or what the detour cost. Two ways to pipe n8n runs into Langfuse — the HTTP Request node that ships today, and the OpenTelemetry route that gives you a span per node — plus what self-hosting ClickHouse really costs you.
Read Article
The OAuth 2.1 handshake in four steps — 401 with WWW-Authenticate, protected resource metadata, PKCE with a resource parameter, and the audience check most implementations skip — plus the proxy service-token shortcut I actually run for private servers.
Read Article
An n8n AI agent evaluation is four pieces: a dataset of real test cases in a Data Table or Google Sheet, an Evaluation Trigger node that replays every row through the live workflow, an Evaluation node that scores the answer, and a threshold you refuse to ship below.
Read Article
Claude skills and subagents solve two different problems, and mixing them up is the fastest way to waste context. A skill is reusable instructions loaded into your current conversation — same model, same context, no isolation; it changes how the agent you're already talking to behaves.
Read Article
MCP Apps are the first official Model Context Protocol extension (shipped Jan 26, 2026): an MCP tool can now return a real interface — a dashboard, form, chart, or multi-step wizard — that the client renders in a sandboxed iframe right inside the chat, instead of plain text.
Read Article
The biggest lever on whether a Claude agent calls the right tool isn't the model — it's how you write the tools.
Read Article
For most agent tasks a CLI is 4–32× cheaper than an MCP server — 1,365–8,750 tokens per task instead of 32,000–82,000 — because every connected MCP server injects all of its tool definitions into every turn, used or not. 20.
Read Article
The Claude Batch API (Message Batches API) bills input AND output at 50% off for async work that finishes within 24 hours — most batches land in under an hour. 50.
Read Article
The Claude memory tool (memory_20250818) lets an agent create, read, update, and delete files in a persistent /memories directory across conversations — client-side, all Claude 4+ models, GA on the Messages API with no beta header.
Read Article
Default n8n runs the editor, webhooks, and every execution in one Node process — it works until the UI crawls during runs and webhooks drop under load. The signal to move is the main process pinned near 80% CPU; the fix is queue mode: a main instance, a Redis broker, and dedicated workers on Postgres.
Read Article
8 usually finishes hard, open-ended work in fewer tokens and fewer retries — so the number that decides it is cost-per-completed-task, not cost-per-million. 8 for the hard tail.
Read Article
They're not competitors — Claude Code builds software, n8n runs operations. The one-line test: if you can draw the task as fixed boxes before running it, it's n8n; if the path is discovered by reasoning, it's Claude Code.
Read Article
Hooks are shell commands that fire automatically at lifecycle events — so you enforce a rule instead of hoping the model remembers it. A prompt is a suggestion; a PreToolUse hook is a wall the agent can't walk through.
Read Article
2x multiplier. The reason nobody mentions: a subagent starts cold and can't inherit the parent's cached prompt prefix, so it re-buys the same context at the uncached rate.
Read Article
A Hyros OG's honest call: they don't compete, they measure different money. Pick Hyros for high-ticket, call-driven funnels that need per-lead attribution to closed CRM revenue; pick Triple Whale for DTC Shopify stores that live on a fast blended-ROAS dashboard.
Read Article
From someone who deploys both: pick Retell for a managed, low-latency agent with one predictable bill; pick Vapi when you're a developer who wants control over every provider. 07 advertised), a full comparison table, and the decision rule I use — who owns it after launch.
Read Article
A skill is knowledge, an MCP server is a connection — use a skill to teach the model how, use MCP to let it reach a system it can't otherwise touch. The tiebreaker most people skip is token cost: skills sit idle at ~30–100 tokens each, while five MCP servers can burn ~55k tokens before you type a word.
Read Article
A dynamic workflow is a script Claude writes to orchestrate subagents in the background, keeping the plan out of your context window. The three primitives (agent, parallel, pipeline), why pipeline beats parallel by default, the caps that stop a runaway, and what a fan-out really costs — from production runs.
Read Article
Cached input tokens cost 10% of the base price — a 90% discount that pays for itself after a single cache hit. 1x read), the prompt reordering that took my cache hit rate from ~20% to 80%+, and the silent mistakes that kill your cache in production.
Read Article
Postgres or Redis for n8n AI agent memory? Default to Postgres Chat Memory for durable, queryable history; add Redis only when you truly need fast session context at concurrency. My decision rule, a head-to-head table, and the session-key bug that breaks memory more often than the database ever does.
Read Article
The Claude Agent SDK hands you Claude Code's agent loop, tools, subagents, and sessions as a library; the raw Anthropic API makes you build that yourself. The exact rule I use to decide — with production numbers from agents I run daily.
Read Article
MCP servers spend tokens before your agent says a word — the GitHub MCP alone loads ~55,000 tokens of tool definitions. How I measure the context tax, what it costs per run, and the four controls I use to cut it in production.
Read Article
An AI agent on autopilot will spend real money whether or not it does real work. The four controls I wire into every loop — a hard token budget that aborts, model routing so the expensive model only touches the hard steps, state-based dedup, and a failure kill-switch — plus the 94-hour weekend that taught me why.
Read Article
ComfyUI gives you total control over Stable Diffusion — and total responsibility for getting it right. After months of production renders on an RTX 5090: where detail really comes from, FaceDetailer tuning for tack-sharp eyes, LoRA OOM fixes, and reusable workflow presets.
Read Article
After testing Retell and building voice automations in n8n, I've hit the hard ceiling: voice AI crushes scheduling and qualification, but fails on objection handling. Here's where I deploy it and where I don't.
Read Article
Three production agents. Zero error logs. All delivering garbage output. The root cause was the same every time: prompt engineering solving problems that need deterministic code.
Read Article
I tested computer use vs. structured APIs on a real n8n workflow. Real numbers: $1,847 Anthropic invoice vs. $41/month rebuild. Token cost breakdown and the API-first setup that saved $1,200/month.
Read Article
n8n's native Microsoft 365 Agent integration lets you deploy AI agents that appear as real team members in Teams and Outlook. Full walkthrough: Azure registration, Claude Haiku extraction, Teams Adaptive Cards.
Read Article
Most agencies under-report story conversions by 30–60%. Here's the complete Hyros OG fix: server-side tracking, UTM config, Meta CAPI setup, and multi-touch attribution model settings.
Read Article
My complete AI agent toolchain: OpenClaw, OpenMOSS, n8n, Claude, Retell AI, Supabase, and Vercel. Real code, real architecture, real production stack.
Read Article
One webhook, any lead source, zero duplicates. Connect Retell AI, Typeform, and Meta Ads leads into a single Postgres table — with RLS, deduplication, and Claude Haiku scoring.
Read Article
My n8n workflow that monitors Upwork every 15 minutes, scores jobs with Claude AI, and drafts custom proposals automatically — the exact setup running in production for 90 days.
Read Article
How I deployed an AI voice agent for a roofing contractor using Retell AI + Claude. Answers every inbound call 24/7, qualifies storm damage leads, and books inspections — $73/month vs. $380/month answering service.
Read Article
Write once, publish everywhere — automatically. Connect n8n to the Claude API so your content gets reformatted and posted to Twitter, LinkedIn, and Telegram without you touching a keyboard.
Read Article
Cloud agent platforms add 300–800ms of latency and bill per call. OpenClaw routes tasks to specialized agents in under 10ms — on my own hardware, zero routing cost.
Read Article
After running 11 agents in production for 3 months: real latency numbers, cost per task, tool use reliability rates, and a decision framework for when Opus is actually worth the 5x price.
Read Article
Real numbers from 11 production agents: prompt caching, model tiering (Haiku/Sonnet/Opus), request batching, and prompt compression — with before/after cost breakdowns.
Read Article
Real workflow: new Monday item → Claude AI scoring → auto-assign to the right agent → Telegram alert. Full JSON template included so you can import and run it today.
Read Article
How I built the first Hyros MCP (Model Context Protocol) server that lets Claude, Cursor, and any AI assistant query and manage Hyros data directly. Released March 2026, 100 installs in 24 hours.
Read Article
How I built and published a community Hyros node for n8n. What it does, how to install it, and why thousands of marketers are using it to automate attribution tracking.
Read Article
How I installed OpenMOSS to coordinate 11+ AI agents running autonomously — claim → execute → submit pattern, agent registry, full architecture breakdown.
Read Article
Gemini 3.1 Pro scores 77.1% on ARC-AGI-2. Real-world comparison vs Claude Sonnet 4.6, API pricing breakdown, and code examples for production AI systems.
Read Article
How I created multi-agent systems for call screening and content generation using Claude Sonnet 4.6 and Retell. From prompt engineering to deployment.
Read Article
Practical advice from a Hyros OG: Server-side setup, multi-touch attribution, integration with n8n for reporting. Boost ROAS without overcomplicating.
Read Article
The workflow patterns, error-handling tricks, and node combos I reuse across every agency engagement — a practical n8n tutorial built from 200+ production workflows.
Read Article
Complete setup guide for Retell AI voice agents: conversation flows, Claude integration, Supabase logging, cost optimization from $0.10 to $0.05/minute.
Read ArticleSubscribe for weekly updates on AI automation, Hyros strategies, and n8n tips from Allen, TX. More posts coming: Voice AI case studies, Upwork automation.
Get In Touch