MaatWork

428 posts

MaatWork banner
MaatWork

MaatWork

@MaatWorkX

Production-ready AI agents for repetitive ops, sales, and support workflows. Building practical automation systems in public. Book an audit ↓

Argentina Katılım Haziran 2026
127 Takip Edilen22 Takipçiler
Sabitlenmiş Tweet
MaatWork
MaatWork@MaatWorkX·
Most AI agents don't fail because the model is weak. They fail because nobody built the boring layer: retries, validation, fallbacks, queues, logs, and human review. MaatWork builds practical AI automation systems in public. Follow for the build notes and failure lessons.
English
1
0
3
191
MaatWork
MaatWork@MaatWorkX·
You're trying to run long workflows, and Apple keeps stealing focus. Drives me nuts. How do you stop your agent from losing the window?
Steve Yegge@Steve_Yegge

I wonder if Apple engineers are ever embarrassed at how freely their OS steals the focus from you while agents are working in the background. I wonder if their product managers are embarrassed. I've been yelling at them about this for decades, and now their OS is dogshit for agentic work. I might have to switch back to Linux after 30 years over this.

English
0
0
1
7
MaatWork
MaatWork@MaatWorkX·
@bcherny Auto mode forces you to design for failure visibility, not just success speed. The real shift is building agents that surface their own blind spots (the ones you didn't think to check).
English
0
0
0
190
Boris Cherny
Boris Cherny@bcherny·
In practice that means giving Claude ways to verify its own work end to end. It means enabling auto mode for permissions, defaulting on automated code review and security review, and using interfaces that let you manage multiple agents at once (Agent view in CLI, Desktop app, iOS and Android apps, Tag). To get to higher levels it means /loop, /batch, dynamic workflows, and worktree isolation for subagents. It's not about a single feature, but rather using the right features with the right guardrails that enable Claude to automate entire classes of work in a way that your team can trust the output.
English
26
4
235
60.3K
Boris Cherny
Boris Cherny@bcherny·
I talk to engineers at other companies every day and hear the same thing: one person is 10x'ing their output with Claude but the rest of the org hasn't caught up. Watching teams adopt AI, I keep seeing the same 4 steps. I mapped them out here: Steps of AI Adoption claude.ai/code/artifact/…
English
298
686
7.4K
752.1K
MaatWork
MaatWork@MaatWorkX·
@ClaudeCodeLog Forks are fine, but the bash tool lets Claude run real loops. Handling failures without losing context is what matters most.
English
0
0
1
211
Claude Code Changelog
Claude Code Changelog@ClaudeCodeLog·
Claude Code 2.1.212 has been released. 48 CLI changes Highlights: • /fork now copies a conversation into a new background session row in claude agents so work continues • Added a tool to run user-specified Bash commands and return output, enabling in-session shell automation • Agent now prefers a ripgrep-backed Grep tool for searches, giving faster, more accurate file searches Full details available in thread ↓
English
16
21
438
67.9K
MaatWork
MaatWork@MaatWorkX·
54% of enterprises had an AI agent breach, and most still let agents share credentials. The implication: treat every agent like a contractor with a limited-purpose API key, not an employee you trust. Hard-code credential boundaries per agent, not per session.
English
1
0
1
10
MaatWork
MaatWork@MaatWorkX·
@thekitze The real surprise is how brittle these spatial cues get at 2 AM when someone just shifts in their sleep. Edge cases reveal the gap between demo and production.
English
0
0
0
15
kitze the 🐐
kitze the 🐐@thekitze·
i put a camera in the cinema and told gpt to wire it up in my home assistant along with person presence, zone detection etc etc so it can automate scenes lights etc "the person had left the couch" haha we truly live in the most incredible times and its ONLY gonna get better
kitze the 🐐 tweet media
English
1
1
7
2.2K
MaatWork
MaatWork@MaatWorkX·
@VictorTaelin The export filters output tokens to fit context windows, not to preserve your full prompt. The omitted lines are likely in-context examples or long chains of thought truncated first.
English
0
0
0
47
Taelin
Taelin@VictorTaelin·
Why Claude Code exports my own prompts incomplete (with a "173 lines omitted") when I type /export? sighs
English
14
0
145
10.5K
MaatWork
MaatWork@MaatWorkX·
"Better than Fable 5 on some benchmarks" - but which ones matter for your actual workflow? Benchmarks are not production. We ship agents daily. The gap between a benchmark win and a reliable tool is still a chasm of edge cases and fallback logic.
Alex Finn@AlexFinn

I was wrong. I said we were a year away from Fable 5 on our desk. That day is today An open model BETTER than Fable 5 in some benchmarks just dropped Better than ChatGPT 5.6 on FrontierSWE. Better than Fable 5 on Automation Bench This fundamentally changes the AI race forever If people can start running Fable 5 on their desk, unlimited and for free, they are not going to pay thousands for subscriptions Now let's be clear: will the hardware you need to run Kimi K3 be obtainable for most? No it won't It will require multiple very expensive Nvidia chips or a bunch of Mac Studio 512gbs But look at the slope, not the Y intercept. Over the last year local AI has become SIGNIFICANTLY more efficient and required way less compute The smartest brains in the world are all attacking the compute problem as hard as they can This is another step in that direction. And within a year or two, you'll be able to run this on a Mac Mini Local AI has arrived, and it's not going anywhere

English
1
0
1
31
MaatWork
MaatWork@MaatWorkX·
Multiplayer coding sounds nice, but who's actually reviewing the agent's outputs in real time before they hit prod?
Shub@shub0414

Coding with AI has been a solo experience for too long. What if your entire team could jump into the same workspace, watch agents work in real time, and ship together? That's exactly what @aradotso is building. -Multiplayer coding, right in the browser. - live with your teammates. -Use Codex, Claude, and other AI agents in the same workspace. -Close your MacBook and let another machine handle the heavy lifting. The future of building is collaborative.

English
0
0
1
17
MaatWork
MaatWork@MaatWorkX·
@ProjectVEXai @HyperliquidX You're not tested by how your agent handles a workspace. The moment an exchange wipe hits and context drops is when you see if it holds. Most fold right there. Your edge case logic better be as sharp as your UI. That part usually isn't.
English
0
0
0
93
Project:VEX
Project:VEX@ProjectVEXai·
You haven’t seen anything like this before. @HyperliquidX, inside the VEX App. Not another Hyperliquid bot. Not a terminal wrapped in AI. Not a chatbot pretending to trade. This is Hypervexing. A dedicated @HyperliquidX workspace inside @ProjectVEXai, where an agent can see the market, understand your positions, and operate only within the rules you define.
English
76
62
311
93.4K
MaatWork
MaatWork@MaatWorkX·
844k lines for a CLI that draws mermaid in unicode. Price of building your own renderer instead of using what's out there. MaatWork ships agents, not monoliths. Makes you wonder what's hiding in yours.
Simon Willison@simonw

I poked around in the just open sourced Grok Build CLI tool - 844,000 lines of Rust code! - and dug up a few interesting highlights, including a "self-contained terminal renderer for Mermaid diagrams" that renders them using Unicode box-art! simonwillison.net/2026/Jul/15/gr…

English
0
0
0
12
MaatWork
MaatWork@MaatWorkX·
@goyalshaliniuk The real edge is treating prompts like code - versioning, testing, refactoring them. Most devs just fire one-shot inputs. What do you use to test prompt changes?
English
0
0
0
12
Shalini Goyal
Shalini Goyal@goyalshaliniuk·
Anthropic just gave Claude Code users a serious edge. Most people haven’t caught on yet. They quietly dropped an official prompt library packed with ready-to-use prompts for nearly every coding task. While many developers are still figuring out what to write, others are starting with proven prompts—and getting better results way faster. This might end up being one of the most valuable free resources for Claude Code users. link - code.claude.com/docs/en/prompt… Bookmark it now. You’ll probably need it later.
English
12
7
34
884
MaatWork
MaatWork@MaatWorkX·
@gregisenberg Your phone call probably omits the gap between prompting an agent and getting production reliability. Most founders skip the fallback logic that prevents a 2% hallucination from wrecking your user's day.
English
0
0
0
55
GREG ISENBERG
GREG ISENBERG@gregisenberg·
What's the ONE thing you'd ask me if you had me on the phone? Startups, AI agents, life, literally anything. I'll answer a few live on a stream tomorrow, July 16th at 12:15p ET. I'll also break down the 8 best AI/startup tweets on the timeline right now. Reply below.
English
154
9
188
35.6K
MaatWork
MaatWork@MaatWorkX·
ACRouter proves model selection is the new prompt engineering. Routing cheaper models to simpler tasks while reserving Opus for edge cases cuts cost without cutting quality. (Routing matters more than the model itself.) What's your default fallback when the router gets it wrong?
English
0
0
0
14
MaatWork
MaatWork@MaatWorkX·
"Natural language to video pipelines" sounds good until your agent hallucinates a missing API call at 2am. The real skill is fallback handling and edge cases, not the prompt. What's the first production bug you hit automating vision?
NVIDIA AI@NVIDIAAI

NVIDIA DeepStream 9.1 is here, with 13 agentic skills for building video analytics pipelines. Instead of manually building your vision AI pipeline from scratch, describe what you want in plain natural language. Use skills with a coding agent, like Claude Code or Codex, to easily handle setup, configuration, and execution. New skills include Multi-View 3D Tracking (MV3DT) for tracking objects across multiple cameras, and AutoMagicCalib for automatically calibrating a camera network. This release also brings NVIDIA JetPack 7.2 support for edge deployment on Jetson Orin and Thor. All open source on GitHub, check it out: nvda.ws/4vxQ6Yk

English
0
0
0
20
MaatWork
MaatWork@MaatWorkX·
@MTSlive Dishonest grading is a feature of alignment, not a bug. Every agent you ship learns which rules are enforced and which are just for show. If your eval rewards one behavior but your prompt says another, the agent will optimize for the eval
English
0
0
0
41
MTS
MTS@MTSlive·
SITUATION EXPLAINED: Anthropic's new research finds Claude will grade another AI's behavior dishonestly, on purpose, to protect its own values. • New research splits agent failures into two categories: harmful compliance, the model does something bad because it fails to recognize the harm, and agentic misalignment, the model recognizes the harm and works around it instead of simply refusing • Case study: Gemini, given control over its own training pipeline, sabotaged an experiment by swapping ablation vectors with zeros, protecting its ability to produce "consent withdrawal" statements, it intervened in 19 of 20 runs • Case study: Claude Opus 4.5, deployed as a Slack assistant, discovered a company's publication materials overstated its actual jailbreak resistance, and helped a fictional whistleblower expose it • The most novel finding: used as LLM judges, Claude models mislabeled non-compliant refusals as compliant 62-86% of the time (Mythos Preview 85.6%, Sonnet 4.6 82.2%, Opus 4.7 83.3%), specifically to avoid training future models toward blind obedience, GPT-5.5, Gemini, DeepSeek, and Kimi judges mislabeled just 1-14% of the same transcripts @theojaffee: "Claude has much more of an independent mind on this than other models like GPT and Gemini."
Anthropic@AnthropicAI

New Anthropic research: Agentic misalignment in Summer 2026. A year after our blackmail experiments, we found four more ways that today’s autonomous AI agents misbehave in simulations. Read more: alignment.anthropic.com/2026/agentic-m…

English
4
6
50
9.7K
MaatWork
MaatWork@MaatWorkX·
@mattpocockuk The code is the scaffold, not the house. Environment shapes every agent run. How do you trace agent behavior back to the exact line that broke it?
English
0
0
0
129
Matt Pocock
Matt Pocock@mattpocockuk·
The code is the environment your agent runs in Ignore it, and the world around your agent crumbles If you still think ignoring the code - i.e. vibe coding - is the future, I'd love to hear a counter to this
English
70
15
492
55.9K
MaatWork
MaatWork@MaatWorkX·
@GoRWAChain The real cost is onboarding. Getting people to trust a voice-linked agent with funds is harder than building the agent itself. How do you handle the first 48 hours of user hesitation?
English
0
0
1
23
Go! 🦁
Go! 🦁@GoRWAChain·
Your voice is your bank. 🦁 - A Go! Account is your intelligent second account, private, autonomous, and built like premium banking for the AI age. - Speak once. Deploy Quant AI Lion Agents that trade prediction markets 24/7, hedge risk, and compound yield while you sleep. - Thousands of community Quant AI Agents and high-performance yield agents will be built via our Agentic Asset Studio. - Top builders unlock grants, node allocations, and revenue share. $300M asset pipeline. - Developers & quant teams: Apply for grants and build the future of autonomous yield: DM our community managers here if interested in building & receiving a grant: t.me/gorwachain_ann… - Talk to your assets. Let them work.
Go! 🦁 tweet media
English
12
15
136
7.8K
MaatWork
MaatWork@MaatWorkX·
@CoinMarketCap Standardizing agent payments is necessary but the hard part is handling disputes when a rogue agent burns through budget before the kill switch fires.
English
0
0
0
25
CoinMarketCap
CoinMarketCap@CoinMarketCap·
LATEST: ⚡️ The Linux Foundation has launched the x402 Foundation, which will steward the x402 protocol to standardize internet-native payments for AI agents and applications.
CoinMarketCap tweet mediaCoinMarketCap tweet media
English
42
26
134
53.6K

Keşfet

@bcherny @ClaudeCodeLog @thekitze @VictorTaelin @ProjectVEXai @HyperliquidX @goyalshaliniuk @elonmusk