标签：#AI

📖 TL;DR GPT-5.6 shipped July 9, 2026 in three tiers Sol (flagship), Terra (balanced), and Luna (cheapest) all tuned for agentic tool calling. All three share a 1M-token context window , 128K max output, and native MCP support in the Responses API. Test any MCP server against Sol, Terra, or Luna in MCP Agent Studio — pick the model, connect a server, and watch each tool call live. OpenAI dropped GPT-5.6 on July 9, 2026 - and this one is aimed squarely at agents. Three models landed at once: Sol , Terra , and Luna . Each is built to call tools, not just chat . That makes testing MCP servers with GPT-5.6 a different exercise than testing a plain chat model. Tool selection is the whole game. I have spent this week pointing all three at MCP servers GitHub, Postgres, Playwright, and multi-server setups. This post is what I learned. You will see which tier to run for which workload , how the new tool-calling features change MCP, and how to test each one free in your browser. Skip it and you will overpay for Sol on jobs Luna handles fine. What Is GPT-5.6? Sol, Terra, and Luna Explained GPT-5.6 is a three-tier model family, not a single model. OpenAI split it by cost and horsepower so you match the model to the job. Here is the lineup, straight from OpenAI's pricing page: Model Built for Input / Output (per 1M) GPT-5.6 Sol Flagship — ambitious agentic work $5.00 / $30.00 GPT-5.6 Terra Balanced — efficient, high-volume work $2.50 / $15.00 GPT-5.6 Luna Fast, affordable — everyday work $1.00 / $6.00 The specs are shared across all three. Every tier gets a 1M-token context window, 128K max output, and a February 16, 2026 knowledge cutoff. So the choice is not about context or capability limits. It is about how much reasoning each task actually needs. New to the protocol these models call? Start with what is Model Context Protocol , then come back. Why GPT-5.6 Changes MCP Tool Calling Here is the part that matters for MCP. GPT-5.6 does not just call tools one at a time it can orc

Here's the thing: speed Test: I Found AI APIs 99% Cheaper Than Premium I have a confession: I've been overpaying for AI APIs for years. Like, embarrassingly overpaying. When I finally sat down and actually benchmarked 15 different models on speed and cost, I couldn't believe what I found. Some of the fastest models out there cost literal pennies per million tokens. Here's the thing — if you're still defaulting to whatever the big labs are pushing, you're leaving serious money on the table. So I spent a week running tests through Global API's infrastructure, hitting endpoints from multiple regions, and crunching numbers until my eyes hurt. What I discovered genuinely surprised me. Check this out: there's a model that pushes 80 tokens per second and costs $0.15 per million output tokens. Compare that to premium options charging $3.00/M and you'll understand why I had to write this down. Let me walk you through everything I found. Why I Even Started This Whole Thing My monthly AI bill got out of control. I'm running a few production apps that do text generation, summarization, and chat, and my December bill made me physically flinch. I knew there had to be faster, cheaper models hiding in the ecosystem — I just hadn't taken the time to actually measure them properly. That's the whole reason I ran these benchmarks. Not for clout, not for content marketing. Pure self-interest. I wanted to know where the actual sweet spots are. Where you get the best speed-per-dollar ratio. Where you can save 70%, 80%, even 99% without tanking your user experience. What I found was honestly kind of shocking. My Testing Setup (For the Nerds) I kept the methodology tight and consistent. Here's exactly how I ran everything: Date: May 20, 2026 Test regions: US East (Ohio) and Asia (Singapore) Prompt: "Explain recursion in 200 words" Output target: ~150 tokens per run Iterations: 10 runs each, averaged the results Streaming: Enabled via SSE Base URL: https://global-apis.com/v1 I measured two k

I gotta say, i Spent a Month Testing Chinese AI APIs — Here's What Actually Wins Look, I'm just an indie hacker trying to ship products without going broke. For the past month I've been obsessively running the four biggest Chinese AI model families — DeepSeek, Qwen, Kimi, and GLM — through every test I could think of. And honestly? I wish someone had given me a breakdown like this before I started. So here's my attempt. No corporate fluff, no hand-wavy "it depends" answers. Just real data from someone who actually pays these bills. Why I Even Started Looking at Chinese Models Honestly, I was a GPT-4o loyalist for the longest time. Then I saw my December API bill and nearly choked. $400+ for what amounted to a few chatbot features and some content generation. That's when a friend told me to check out DeepSeek and Qwen. I was skeptical. Like, REALLY skeptical. Chinese models in 2023 were a joke for English tasks. But I kept hearing whispers from other indie hackers about how good things had gotten. So I decided to actually test them properly through Global API's unified endpoint (more on that later). What I found kinda blew my mind. The Quick Cheat Sheet Here's the TL;DR table I wish existed when I started. I'm putting it up top because, lets be real, you probably just want the bottom line: Feature DeepSeek Qwen Kimi GLM Developer DeepSeek (幻方) Alibaba (阿里) Moonshot AI (月之暗面) Zhipu AI (智谱) Price Range $0.25-$2.50/M $0.01-$3.20/M $3.00-$3.50/M $0.01-$1.92/M Best Budget Pick V4 Flash @ $0.25/M Qwen3-8B @ $0.01/M N/A GLM-4-9B @ $0.01/M Best Overall V4 Flash @ $0.25/M Qwen3-32B @ $0.28/M K2.5 @ $3.00/M GLM-5 @ $1.92/M Code Generation ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐⭐⭐ Chinese Language ⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐⭐ English Language ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐⭐⭐⭐ Reasoning ⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐ Speed ⭐⭐⭐⭐⭐ ⭐⭐⭐⭐ ⭐⭐⭐ ⭐⭐⭐⭐ Vision/Multimodal Limited ✅ (VL, Omni) ❌ ✅ (GLM-4.6V) Context Window Up to 128K Up to 128K Up to 128K Up to 128K API Compatibility OpenAI ✅ OpenAI ✅ OpenAI ✅ OpenAI ✅ Alright, now let me act

Teams make dozens of architectural decisions every month but document almost none of them. The rest dissolve into Slack threads, hallway conversations, and the minds of people who will leave the company within a year. Six months later, a new developer stares at the code and asks: "Why Redis here instead of PostgreSQL for queues?" Nobody remembers. An archaeological dig through Git history, Slack, and Notion begins. Two hours spent investigating a decision that originally took 15 minutes. Architecture Decision Records (ADRs) solve this problem. But they don't get written. The reason is simple: drafting an ADR takes 30-40 minutes, and the developer has already moved on to the next task. AI compresses that to 3-5 minutes. This article covers ADR structure, prompts for LLM-based generation, real-world examples, and CI pipeline automation. What ADRs are and why capturing architectural decisions matters An ADR (Architecture Decision Record) is a document that captures one specific architectural decision. Not a spec, not an RFC, not a design document. One decision, one file. Michael Nygard introduced the concept in 2011. The format took hold at large companies (Spotify, Thoughtworks, GitHub) but remains rare in smaller teams. The main reason: the writing overhead feels higher than the value it delivers. Three situations where the absence of ADRs hurts the most: Onboarding. A new developer reads the code and encounters an unconventional decision. Without an ADR, they either spend hours investigating, or treat it as a mistake and "fix" it. Both paths are expensive for the team. Revisiting decisions. Context changes: load increases, new requirements emerge, a dependency goes stale. Without a record of why the current solution was chosen and which alternatives were rejected, the team re-runs the entire analysis from scratch. Audits and compliance. In regulated industries (fintech, healthtech), architectural decisions require documented justification. ADRs close that gap automa

I spent the better part of the last decade writing enough boilerplate code to regret it. In the early PHP days, it was FTPing files; in the modern era, it's writing custom Python scripts just to check if a new Claude model is out or to see if my prompt is going to blow my budget on tokens. We have reached a point where we are building 'agentic workflows,' yet the first thing every developer does when they want an agent to interact with Anthropic is write an API wrapper. It's redundant work. If you're using Claude in Cursor or Claude Desktop, the model should be able to talk to its own source. The Anthropic MCP server changes this by turning the Messages API into a set of tools rather than a separate integration task. It turns your AI agent into an orchestration layer for the API itself. The problem with 'Just use the API' When you're building with LLMs, there's a hidden tax: context management and cost uncertainty. You send a prompt, it works. You send a slightly larger one, it hits a context limit or costs three times what you expected. If your agent has access to the count_tokens tool via MCP, the workflow changes fundamentally. Instead of blindly sending massive payloads and praying to the provider gods, the agent can 'pre-flight' a prompt. It can look at the messages array, calculate the input token count, and decide—without human intervention—whether it needs to truncate context or if it's safe to proceed. This isn't just about convenience; it's about building reliable, autonomous systems that don't fail halfway through a complex reasoning task because they hit a hard limit. Managing the heavy lifting: Batching as a first-class citizen The most underrated tool in this set is create_batch_message . If you've worked with Anthropic's batch API, you know it’s the only way to handle high-volume, independent requests without destroying your budget. It's 50% cheaper than standard requests. But managing batches traditionally is a pain in the neck. You have to submit th

I kept re-running the exact same input through my translation app. Same code. Same model. Same everything. And the word "machines" kept flipping between two different translations. Sometimes it came out as "機械" (machine). Sometimes as "あなたのPC" (your PC). No code changed between runs. No input changed either. My first assumption was a race condition somewhere in my pipeline. It wasn't. Where I actually looked I checked the obvious suspects first: caching, threading, anything stateful that could make the same input behave differently on different runs. All clean. So I went one level deeper, into how the model picks the winning word. Translation models score every candidate word and pick whichever scores highest. When I logged the actual scores for "machine" vs "your PC" on this input, they were almost exactly tied. That's the part that mattered. When two candidates are separated by a tiny margin, the order floating-point operations get summed in can nudge the score just enough to flip which one wins. Same math, same inputs, different accumulation order between runs — and a near-tie flips sides. Nothing was actually random. It was deterministic all the way down. It just wasn't deterministic in a way I could predict, because the thing that decided the winner was rounding noise several layers below anything I was testing. The fix wasn't "make it deterministic" Forcing strict floating-point determinism across an ML pipeline is its own rabbit hole, and not one I wanted to go down for one word. Instead, I looked at why the tie was so close in the first place. "Machine" and "your PC" were close enough in meaning, in this context, that the model wasn't confident either way. So I widened the margin instead of trying to eliminate the noise: I swapped the input word choice from "machines" to "equipment," which the model was much more decisively confident about. Scores stopped being close enough for rounding noise to matter. The flip-flopping stopped. I want to be honest about a

They're rolling up their sleeves again, seemingly out of fear of missing AI's defining moment and, presumably, the irresistible allure of making even more money -- potentially a lot more.

AudioTrust: reconciliar C2PA y watermark AudioSeal en audio sintético Un verificador local que lee las dos marcas de confianza de un audio generado por IA (procedencia C2PA + watermark AudioSeal) y emite un veredicto auditable sobre si coinciden, se contradicen o faltan. El problema Un audio sintético puede llevar dos marcas de confianza distintas: Procedencia C2PA : un certificado digital embebido en el archivo (su "DNI" de origen — quién, cuándo, con qué herramienta). Watermark AudioSeal : un código inaudible incrustado en el sonido, detectable aunque el audio se comparta o transcodifique. Cada una por separado es útil, pero ninguna es suficiente. La procedencia puede faltar (mucho audio generado no la incluye) y el watermark puede estar presente en audio totalmente legítimo. El caso interesante es cuando se contradicen : el manifest C2PA dice "grabado por un humano con una grabadora" pero el watermark de una herramienta de IA está presente. Eso es una señal de manipulación — el llamado Integrity Clash . AudioTrust no genera ni firma nada. Es un verificador : lee ambas capas y las reconcilia. Qué hace audio.wav ──► AudioTrust verify ──► veredicto + explicación C2PA watermark Veredicto ausente ausente unverifiable ausente presente partial origen sintético presente trusted origen humano presente contradiction (Integrity Clash) Salida JSON: { "file" : "audio.wav" , "verdict" : "trusted" , "c2pa" : { "present" : true , "source_type" : null , "claims" : [ "action=c2pa.created by TestTTS" , "generatedBy=TestTTS" ]}, "watermark" : { "present" : true , "detect_prob" : 0.92 }, "explanation" : "C2PA declara origen sintético y hay watermark fuerte: coherentes." } Cómo funciona Lectura C2PA con c2pa-python (el Reader de la librería oficial). Si no hay manifest, devuelve present=False sin crashear. Detección de watermark con audioseal . Devuelve solo detect_prob (P(audio watermarked) en [0,1]). Reconciliación determinista en reconcile.py . Dos decisiones de diseño que vale la

Here is a reliable test: find a prompt that isn't working. Read it carefully. Now ask yourself — at which specific sentence did the model get permission to do what it did wrong? You will almost always find it. A hedged instruction. A missing constraint. An ambiguous scope. The model did not misunderstand you — it followed the most statistically probable interpretation of what you wrote. That interpretation was not the one you intended. These are not beginner mistakes. They are structural patterns that reappear at every experience level, because they look reasonable when you write them and only reveal themselves in the output. TL;DR: Prompts fail because they hand interpretive control to the model on dimensions where you had a specific requirement. Each of the seven mistakes below is a different way of doing that — and each has a specific, testable fix. Mistake 1: Placing Critical Instructions in the Middle of the Prompt Language models process all tokens simultaneously through attention mechanisms , but the effective weight any individual token receives depends heavily on its position. Instructions near the beginning and end of a prompt receive disproportionately more attention weight than those in the middle. This is not a quirk — it is a consequence of how positional embeddings interact with self-attention across long contexts. This effect is well-documented. The "Lost in the Middle" study (Stanford / UC Berkeley, 2023) showed that retrieval accuracy from long-context windows degrades significantly for information placed in the middle — even in capable models. The same mechanism applies to instruction prompts: GPT-4o and Claude 3.5 Sonnet both exhibit measurably lower constraint adherence for instructions buried mid-context compared to those at the leading or trailing position. Open-weight models including DeepSeek-V3 and Llama 3 display the same positional bias — this is not a proprietary model quirk, it is a structural property of the transformer architecture. T

Uber Chief Product Officer Sachin Kansal walks TechCrunch through the company's financial-services ambitions, its increasingly complicated relationship with Waymo, its new AV Labs data operation, and how AI is starting to show up in ways riders and drivers will actually notice.

In my last post I said that for normal HTML pages, element-based automation ( find / read_page ) beats coordinates every time. This post is about the apps where that advice is useless. Flutter Web apps. Canvas-rendered editors. Every button and panel you can see on screen doesn't exist in the DOM — it's all pixels painted onto a single canvas. find returns nothing. read_page 's accessibility tree is effectively empty. I got Claude to drive the Rive editor (an animation tool built with Flutter) all the way through selecting assets and exporting them. Here's the procedure that survived contact with reality. Step zero: confirm you're actually in this situation Coordinate automation is fragile, so you should only accept it after ruling out the alternative. The test is quick: run read_page . If the visible UI has almost no corresponding nodes, you're looking at a canvas-rendered app, and coordinates are the only interface you have. The four rules 1. Wait for the window size to settle before anything else Same failure mode as my previous post: right after load, the viewport hasn't reached its final width (I measured 1664→1920 over 2–3 seconds), and clicks based on an early screenshot land to the right of the target. Read innerWidth via javascript_tool twice; only proceed when two consecutive reads match. But matching innerWidth alone isn't enough — also confirm devicePixelRatio hasn't changed since the screenshot you're about to act on (a follow-up to my previous post surfaced this: when DPI or scaling changes, the whole coordinate space rescales the same way, but the new values stabilize immediately, so an innerWidth -only check can't catch it). Canvas apps deserve extra paranoia here, because there is no element-based fallback when a click misses. 2. Read text by zooming, not by extracting Text painted on canvas can't be pulled out of the DOM. To read a menu item or panel label, zoom into that region and read the enlarged screenshot as an image. Full-page screenshots ma

I Built an AI-Powered CLI That Migrates Legacy Java Code to Java 17/21/25 If you've spent any time in enterprise Java, you know the feeling. You open a service that's been running since 2014 and you're greeted by walls of anonymous inner classes, verbose null checks, Collections.unmodifiableList wrapping a new ArrayList , and switch statements with more break keywords than actual logic. Individually each pattern takes 30 seconds to fix. But across a codebase with 300 files, it's a week of mechanical work — and that's before you factor in the code review. So I built java-migrate : a CLI tool that scans your Java files, detects legacy patterns, and sends them to Claude with a precise system prompt to get them modernised. One command, instant diff, no surprises. What it looks like in practice Here's a typical legacy file before migration: public class LegacyService { // POJO with getters/setters public static class User { private String name ; private int age ; public String getName () { return name ; } public void setName ( String name ) { this . name = name ; } public int getAge () { return age ; } public void setAge ( int age ) { this . age = age ; } } public List < User > sortUsers ( List < User > users ) { // Anonymous Comparator users . sort ( new Comparator < User >() { @Override public int compare ( User a , User b ) { return a . getName (). compareTo ( b . getName ()); } }); return Collections . unmodifiableList ( users ); } public String describeObject ( Object obj ) { // instanceof + cast if ( obj instanceof String ) { String s = ( String ) obj ; return "String of length " + s . length (); } return "Unknown" ; } public String classify ( int value ) { // switch statement String result ; switch ( value ) { case 1 : result = "one" ; break ; case 2 : result = "two" ; break ; default : result = "other" ; } return result ; } } Run java-migrate LegacyService.java --dry-run --verbose and you get this diff: - public static class User { - private String name; - privat

Every time I shipped something, the same thought hit me a few hours later: It works on my machine. It works in staging. But is it actually working for the people using it right now? I had analytics. I had a green dashboard. And I still had no honest answer to that question. Users would quietly leave, a button would silently break on Safari, a page would crawl on a mid-range Android, and I'd find out days later, if at all. That gap is what I ended up building HeronSignal to close. But before I talk about the tool, let me talk about the pain, because I think you've felt at least one version of it. The pain, depending on who you are If you're a vibe coder / solo builder You ship fast. Cursor, Claude, v0, a Vercel deploy, and it's live. Beautiful. Then… nothing. You have no idea what happens after "Deploy successful." Is the checkout button throwing an error on mobile? Is your landing page slow enough that half your visitors bounce before it paints? You don't know, because setting up "real" monitoring feels like a second job: a Datadog dashboard you'll never look at, a Sentry config you half-finish. So you just… hope. And hope is not a monitoring strategy. If you're an engineer Your problem isn't no data. It's too much . Ten dashboards, alert fatigue, a Sentry inbox with 400 issues where 390 are noise. Something's clearly wrong, but which thing actually matters? You spend your morning triaging instead of fixing. And when you finally pick an error, you get a stack trace with zero context: no idea what page it happened on, what the user was doing, or how to reproduce it. Triage is not the job. Fixing is the job. But the tools make you do the triage first. If you're a product person You can see in your funnel that people drop off at step 3. What you can't see is why . Was it a JS error? A slow page? A confusing layout? Your analytics tool tells you what happened but never why , and the engineering dashboards that might explain it are unreadable walls of numbers. So you gue

As developers, we all want AI to review our code. But sending proprietary, unreleased code to third-party cloud APIs (like OpenAI or Anthropic) isn't always an option—especially if you're working on client projects or under strict NDAs. I wanted an AI code reviewer that was 100% private , free , and actually understood the context of my entire project . So, I built one using Python and Ollama . Here’s a look at what it does and how you can use it! What it does It’s a CLI tool that uses local LLMs (like qwen2.5-coder or llama3 ) to review your code. No API keys, no subscriptions, and zero data leaves your machine. But I didn't want to just paste code snippets into a terminal. I wanted a tool that actually fits into a developer's workflow. Here is what it supports: 1. Review an Entire Codebase Just point it at your project folder. The app will recursively gather your files, automatically ignoring bulky folders like node_modules , .git , vendor , and .next , and give you a full architectural review. python3 app.py ./my-project/ 2. Review Pull Requests Automatically Want to review a PR? Just pass the GitHub PR URL. The tool auto-detects that it's a diff, fetches the changes, and switches into "PR Review Mode." Instead of looking at architecture, it zeroes in on the + lines to find bugs, edge cases, and missing tests introduced by the PR. python3 app.py https://github.com/facebook/react/pull/30000 (Working on a private repo? Just pipe it: gh pr diff 123 | python3 app.py ) 3. Pipe Anything Into It You can pipe individual files, diffs, or snippets straight from your terminal. cat src/main.py | python3 app.py 🛠️ How to run it yourself Install Ollama and pull a solid coding model: ollama pull qwen2.5-coder Clone the repo and install the requirements: pip install -r requirements.txt Run it! python3 app.py ./your-code 💡 The Magic Under the Hood The script dynamically switches its prompt based on what you feed it. If you give it a directory, it looks for separation of concerns

The AI hardware landscape has shifted significantly in 2026, with NVIDIA, AMD, and Intel all competing for developers who need GPUs capable of running local large language models and AI inference workloads. Choosing the right GPU for AI workloads requires looking beyond marketing numbers and focusing on the specifications that actually affect real-world performance. Memory capacity, memory bandwidth, and software ecosystem maturity consistently matter more than theoretical compute peaks when running transformer models locally. This comparison covers the most relevant workstation and prosumer GPUs available in mid-2026, including NVIDIA's Blackwell architecture (RTX 50-series), AMD's Radeon AI Pro R9700, and Intel's Arc Pro B70. The goal is to provide a practical reference for developers deciding which hardware best fits their model sizes, software stack, and budget constraints. Which GPU specifications matter for AI workloads Marketing materials from GPU vendors emphasise AI TOPS and tensor performance, but these metrics rarely tell the complete story for local inference. The specifications below are ranked by their actual impact on running large language models. VRAM capacity VRAM is typically the first limiting factor when running LLMs locally. A model cannot execute entirely on the GPU if it does not fit into available memory. Once model weights spill into system RAM, inference performance drops dramatically. Approximate VRAM requirements for common model sizes: Model Size Recommended VRAM 7B 8-12 GB 14B 16 GB 32B 24-32 GB 70B 48-64 GB 120B+ Multiple GPUs For most homelab users, moving from 16 GB to 32 GB of VRAM provides a substantially larger practical benefit than increasing raw compute performance. A 32 GB GPU capable of running an entire model will often outperform a theoretically faster 16 GB GPU forced to offload tensors into system memory. Memory bandwidth Memory bandwidth determines how quickly model weights can be streamed into compute units. Large tran

Your customers have stopped Googling "best note-taking app." They're asking ChatGPT, Perplexity, and Gemini instead — and getting back a short list of three or four products. If your brand isn't on that list, you're invisible, and unlike a Google ranking you can't even see where you stand. That's the problem I set out to measure. This post is the build breakdown: five AI answer engines, one uniform result shape, a mention-detection core that doesn't lie to you, and the honest gotchas I hit around cost and billing. The whole thing runs as a paid Apify Actor written in async Python. The niche has a name now — GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization). Think SEO, but the search engine is a language model and the "ranking" is whether you get named in the answer. The core question Give the tool a brand, its competitors, and the buyer-intent questions your customers actually type: { "brand" : "Notion" , "competitors" : [ "Obsidian" , "Coda" , "Evernote" ], "prompts" : [ "best note taking app for students" , "Notion vs Obsidian which should I use" ], "engines" : [ "perplexity" , "chatgpt" , "gemini" , "claude" , "aiOverview" ], "samplesPerPrompt" : 3 } It asks each engine each prompt (several times, because LLM answers vary run-to-run), then analyzes every answer for: were you mentioned, how early, were you recommended or just listed, what's the sentiment, who else got named, and — the part incumbents skip — which domains each engine cited. That last one is the actionable output: it tells you which websites the AI trusts for your category, i.e. where you need coverage. Architecture: one shape to rule them all The trick that keeps the whole thing sane is that every engine adapter — whether it's a clean REST API or a messy HTML scrape — returns the exact same record shape : { " engine " : " perplexity " , " prompt " : " best note taking app for students " , " sampleIndex " : 1 , " responseText " : " ... " , " citations " : [{ " url " : " ... "

With the cash, the company aims to expand its world model offering and reach customers across geographies.

The company is raising at least $75 million, led by Robot Ventures, with significant participation from USV and other prominent investors.

When an autonomous agent gets an email address of its own, the first question your security team asks isn't "can it send mail?" It's "can you prove, six months from now, exactly what it said and to whom?" That's a different problem from "does it work." A demo that fires off a few support replies looks great in a sprint review. But the moment a real customer says "your bot promised me a refund," or a regulator asks for the complete record of what an automated system told a data subject, you need a defensible trail — an immutable record of every outbound and inbound message the agent touched, captured outside the mailbox the agent can also delete from. I work on the Nylas CLI, so the terminal commands below are the exact ones I reach for. But the architectural point here is provider-agnostic and it's the part most "AI email" tutorials skip: the live mailbox is not your audit log. It's mutable, it has retention limits, and the same agent that sends mail can also trash it. If your only record of what the agent did lives in the inbox, you don't have an audit trail — you have a working copy. What "audit-log everything" actually means There are two stores in this design, and keeping them separate is the whole point. The live mailbox — the Agent Account grant. Messages flow in and out here. It's queryable, it's real-time, and it's mutable . Flags change, messages move folders, things get trashed. On the free plan it's also retention-limited: 30 days for the inbox, 7 days for spam. The audit store — your system. An append-only, write-once log keyed by message_id and thread_id . Nothing in it is ever updated or deleted in normal operation. This is the record you hand a reviewer. The audit store is the thing you build. Nylas gives you the two capture points — the send response and the inbound webhook — but the immutability is your responsibility. That means a WORM (write-once-read-many) object store, an append-only table with no UPDATE / DELETE grant for the app role, or a has

Most multi-tenant SaaS apps that send email do it from one shared identity. There's a notifications@yourapp.com , every customer's mail flows through it, and the tenant is just a from_name you stamp on the subject line or a footer you swap out. That's fine until it isn't — until Tenant A's spam complaints drag down Tenant B's deliverability, until a reply from a customer lands in a single firehose inbox you now have to fan back out, until one tenant wants a stricter send cap than another and you realize you built none of that into the data model. So let's not share. Let's give every tenant its own real mailbox — a dedicated Agent Account per customer, each with its own grant_id , its own send identity, its own policy and limits, grouped into its own workspace. Not one inbox with a thousand label hacks. A thousand inboxes, isolated by construction. I work on the Nylas CLI, so the terminal commands below are the exact ones I reach for when I'm wiring this up. Every step gets the two-angle tour: the raw curl call and the nylas command that does the same thing. Why per-tenant beats one shared sender The shared-sender model fails along a few predictable seams. Per-tenant Agent Accounts close each one: Deliverability blast radius. When everyone sends from one address, one tenant's bounce rate and spam complaints poison the reputation everyone shares. Per-tenant accounts — and, if you want, per-tenant domains — keep one customer's bad behavior from sinking the rest. Inbound that actually belongs to someone. A shared sender means replies come back to one mailbox and you're left correlating them to tenants by hand. When each tenant has its own grant, an inbound message.created event already carries the grant_id . The routing is done before your handler runs. Per-tenant policy and limits. Different customers, different rules. A trial tenant capped at a low daily send; an enterprise tenant with a higher quota and longer retention. With a shared sender you'd build all of that y