The perfect hat doesn’t exi-

@itamar_mar @QodoAI @MiniMax__AI @simpsoka @GoogleLabs @NM_Financial @business @jedimody @browsercompany @danshipper @every Okay, and that’s a wrap on the leadership track day! 🙌🏽 thanks to @swyx for organizing an incredible summit!!

@itamar_mar @QodoAI @MiniMax__AI @simpsoka @GoogleLabs @NM_Financial @business @jedimody @browsercompany 14/ “AI-Native Company Building” - @danshipper (@every) Dan showed what an AI-native org looks like: four products, 7k paid subs, 100k free users, and 99% of code written by agents with only $1M raised and just 15 people. 🤯🤯

@itamar_mar @QodoAI @MiniMax__AI @simpsoka @GoogleLabs @NM_Financial @business 13/ “Dia: Building an AI-Native Browser” - @jedimody (@browsercompany) Dia talked through a Jepa-inspired loop for better model behavior: mutate prompts, score them, reflect, and keep iterating. It gives them a fast path to refinement without retraining or heavy infra.

@itamar_mar @QodoAI @MiniMax__AI @simpsoka @GoogleLabs @NM_Financial 12/ “Deploying AI within Bloomberg’s Eng Org” – Lei Zhang (@business) Bloomberg used agents for patches, migrations, and incidents, not just codegen. Wins came from MCP-powered context, unified tooling, and embedding AI in onboarding so new devs spread adoption across teams.

@itamar_mar @QodoAI @MiniMax__AI @simpsoka @GoogleLabs 11/ “Small Bets, Big Impact” - Asaf Bord (@NM_Financial) Northwestern Mutual built GenBI to answer data questions like a BI partner. A 4-phase rollout delivered value each sprint, automating report discovery and metadata work while keeping enterprise trust and controls.

@itamar_mar @QodoAI @MiniMax__AI 10/ “Proactive Agents” - @simpsoka (@GoogleLabs) Google showed Jules, a proactive coding agent that cuts dev mental load. Instead of waiting for prompts, it anticipates needs: fixing issues, learning style, offering context, and soon understanding consequences across the stack.

@itamar_mar @QodoAI 9/ “MiniMax M2” - Olive Song (@MiniMax__AI) M2 is a 10B open-weight model built for coding agents. It’s cost-efficient, ranks top in intelligence and agentic benchmarks, and is already #3 on OpenRouter by tokens. Strong real-world usage suggests traction far beyond benchmarks.

8/ “The State of AI Code Quality: Hype vs Reality” - @itamar_mar (@QodoAI) AI code use is soaring (82-92% of devs), but so are bugs. Same bug rate per line, far more lines. Needed: AI-powered quality gates, stronger context engines, and automated validation across the SDLC.

7/ “How to Quantify AI ROI in Software Engineering” - Yegor Denisov-Blanch (@Stanford) A 120k-developer study showed AI speeds coding but overwhelms review and adds tech debt. Real ROI came when teams paired generation with AI-led testing and review.

6/ “Moving Away from Agile” - Martin Harrysson & @maniarnatasha (@McKinsey) Highlighted the AI productivity gap: big individual gains but only 5–15% org impact. Their answer is an AI-native model with 3–5 person pods, continuous planning, and spec-driven development.

5/ “Future-Proof Coding Agents” - @realchillben & @bfioca (@OpenAI) The real frontier for coding agents isn’t the UI or the model but the harness: the prompts + tools wrapped around it. Models shift constantly; a stable harness lets teams upgrade w/out breaking agent workflows.

Turns out I love robot dogs too 😂

4/ “2026: The Year the IDE Died” - @Steve_Yegge (@Sourcegraph) & @RealGeneKim Traditional IDEs won’t survive the agent era: there’s already a ~10x gap between engineers who use agents vs those who don’t. Future dev looks like swarms of specialized agents, not bigger IDEs.

3/ “Your Support Team Should Ship Code” - Lisa Orr (@zapier) Lisa likened Zapier’s 8,000+ apps to Grand Canyon layers and support to the river spotting erosion. Their agent Scout now triages issues, checks fixability, writes MRs, and powers ~40% of fixes.

2/ “Autonomy Is All You Need” - @pirroh (@Replit) We need to empower the non-technical user. Today's agents are Tesla: powerful but still needing a “driver.” The goal is Waymo-level autonomy for software creation, powered by strict verification and agent-led task decomposition.

1/ Evolving Claude APIs for Agents - @katelyn_lesse (@AnthropicAI) Claude Code is becoming a real agent platform: reliable tool use, built-in search, and custom tools with proper schemas. The real unlock is context: MCP + memory + context editing gave them a ~39% perf boost.

0/ Reporting live(ish) from the @aiDotEngineer summit 🫡 - here are the most interesting things I heard this morning across @AnthropicAI, @Replit, @zapier, and @Sourcegraph / @AmpCode. 🧵↓

Let’s go @aiDotEngineer Day 1! So fun to have a mini @DecibelVC portco meetup with @joshdevonai & @mjamei + after-party with the @sondera_ai team 🔥 🍺

my new drinking game is a terms bingo card for AI pitches. squares include “ontology,” “orchestration layer,” “model router,” and “context engineering.” which am I missing 😂

Too real