Fostering human connection with OurStories (Stanford/Cornell CS, Meta, Trulia/Zillow, Netscape/AOL). I love teaching AI, mentoring tech startups, and building.
We are entering an age of effortless creation. Yes, that means AI created content (including AI slop) will proliferate. But my hope is that human created digital content will also increase (assisted by tools that lower the friction of creation). I also hope that we end up creating more content (human or AI generated) that fosters human to human connection, understanding, and empathy.
Are you a software engineer who wants to become an AI Engineer? Read on, and please share. The "๐๐ ๐๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ ๐๐ฐ๐ฎ๐ฑ๐ฒ๐บ๐" series offers milestones and live discussions to guide our journey to becoming proficient AI engineers.
๐ ๐๐ ๐๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ ๐๐ฐ๐ฎ๐ฑ๐ฒ๐บ๐, ๐๐ผ๐ต๐ผ๐ฟ๐ #๐ญ, ๐ฃ๐ผ๐๐ #๐ฑ
๐ ๐ ๐ถ๐น๐ฒ๐๐๐ผ๐ป๐ฒ ๐ฑ: ๐๐ฎ๐ฟ๐ป๐ฒ๐๐ ๐๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ๐ถ๐ป๐ด & ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐๐ผ๐ผ๐ฝ๐: ๐ฆ๐ฐ๐ฎ๐ณ๐ณ๐ผ๐น๐ฑ๐ถ๐ป๐ด ๐ณ๐ผ๐ฟ ๐๐ผ๐ป๐ด-๐ฅ๐๐ป๐ป๐ถ๐ป๐ด ๐๐ด๐ฒ๐ป๐๐
In Milestone 4, we learned to build, evaluate, and connect agents. But a model alone isn't enough. The harness (prompts, tools, memory, and control logic) and the loop that drives the agent over long horizons matter just as much. This milestone covers that scaffolding.
โ ๐ฆ๐๐ฒ๐ฝ ๐ญ) ๐๐ผ๐๐ป๐ฑ๐ฎ๐๐ถ๐ผ๐ป: ๐ช๐ต๐ฎ๐ ๐๐ ๐ฎ ๐๐ฎ๐ฟ๐ป๐ฒ๐๐ ๐ฎ๐ป๐ฑ ๐ช๐ต๐ ๐๐ผ๐ป๐ด-๐ฅ๐๐ป๐ป๐ถ๐ป๐ด ๐๐ด๐ฒ๐ป๐๐ ๐ก๐ฒ๐ฒ๐ฑ ๐ข๐ป๐ฒ: Read Anthropic's "Effective Harnesses for Long Running Agents" and its companion "Harness Design for Long Running Apps." They explain why agents that run for minutes or hours, not seconds, need deliberate engineering for context compaction, error recovery, and state management.
โ ๐ฆ๐๐ฒ๐ฝ ๐ฎ) ๐๐ฎ๐ฟ๐ป๐ฒ๐๐ ๐๐ฒ๐๐ถ๐ด๐ป ๐ถ๐ป ๐ฃ๐ฟ๐ฎ๐ฐ๐๐ถ๐ฐ๐ฒ: Read OpenAI's "Harness Engineering: Leveraging Codex in an Agent-First World" for a practical look at building harnesses around coding agents in a real agent-first workflow.
โ ๐ฆ๐๐ฒ๐ฝ ๐ฏ) ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐๐ผ๐ผ๐ฝ๐: ๐ฅ๐ฒ๐น๐ถ๐ฎ๐ฏ๐น๐ฒ ๐ช๐ผ๐ฟ๐ธ๐ณ๐น๐ผ๐๐ ๐ง๐ต๐ฎ๐ ๐๐ฒ๐ฎ๐ฟ๐ป ๐ข๐๐ฒ๐ฟ ๐ง๐ถ๐บ๐ฒ: Every agent runs on a loop: observe, decide, act, repeat. "How to Solve Problems of Autonomous Agentic AI Workflows" covers common failure modes (drift, runaway loops, lost context) and how to keep a loop on track. LangChain's "Continual Learning for AI Agents" shows how the best loops learn from each iteration and improve with every run.
โ ๐ฆ๐๐ฒ๐ฝ ๐ฐ) ๐ฃ๐๐๐๐ถ๐ป๐ด ๐๐ ๐ง๐ผ๐ด๐ฒ๐๐ต๐ฒ๐ฟ: ๐๐ป ๐๐น๐๐ฎ๐๐-๐ข๐ป ๐๐ด๐ฒ๐ป๐ ๐ฌ๐ผ๐ ๐๐ฎ๐ป ๐ ๐ฒ๐๐๐ฎ๐ด๐ฒ: Study the "Hermes Agent Masterclass" (architecture similar to OpenClaw): the harness and loop working together as an always-on agent that lives on a server, runs continuously, and reaches you over Telegram or Slack. Send it a task from your phone, and it works on its own and messages you back.
Do the above and comment below with your questions.
๐ ๐๐ถ๐๐ฒ ๐๐ ๐: ๐ช๐ฒ๐ฑ๐ป๐ฒ๐๐ฑ๐ฎ๐, ๐๐๐น๐ ๐ญ, ๐ต:๐ฏ๐ฌ ๐ฎ๐บ ๐ฃ๐ฆ๐ง Comment below for the Zoom link and bring your questions.
๐ ๐๐ถ๐ป๐ธ๐ (๐ณ๐ผ๐ฟ ๐๐๐ฒ๐ฝ๐ ๐ฎ๐ฏ๐ผ๐๐ฒ):
๐ฆ๐๐ฒ๐ฝ ๐ญ ๐๐ถ๐ป๐ธ๐: 1) https://t.co/TVsZjp0f2J, 2) https://t.co/TXvnF2ylXF
๐ฆ๐๐ฒ๐ฝ ๐ฎ ๐๐ถ๐ป๐ธ: https://t.co/kJsdFrAb1q
๐ฆ๐๐ฒ๐ฝ ๐ฏ ๐๐ถ๐ป๐ธ๐: 1) https://t.co/BiH0uTNnOg, 2) https://t.co/q9YW68j2Xo
๐ฆ๐๐ฒ๐ฝ ๐ฐ ๐๐ถ๐ป๐ธ: https://t.co/ekRgITZJ7h
Kudos to @kieranklaassen of @every on writing this Compound Engineering guide: https://t.co/lhlozbPdjF. Highly recommend that everyone building with AI read this guide.
Autonomous companies will become a reality before the end of 2026.
I want autonomous companies to be accessible for anyone with a computer, a dream, and Pancake.
Not just the top 0.1%.
Introducing Pancake: the OpenClaw cofounder that makes your company autonomous.
I constantly get asked about quantization by students I teach in agentic AI courses. It is a technique used to compress model weights so that model inference can run cheaper and faster. @samwhoo at @ngrokHQ has written a fantastic post: "Quantization from the ground up" that explains everything from first principles. Highly recommend! https://t.co/n7Tcl03bZr
Here's a Ralph loop based plan/build workflow that is working well for me:
1) Talk to Claude Code about new features. Generate multiple options with pros/cons. Then ask CC to write a .md spec in specs/ directory.
2) Run Ralph loop in plan mode to create an IMPLEMENTATION_PLAN.md. This usually includes other architectural and platform improvement suggestions not all of which I want to implement immediately. Edit this plan to my satisfaction.
3) Run Ralph loop in build mode for an appropriate number of iterations to implement everything in IMPLEMENTATION_PLAN.md. In every loop, 1-2 work items from the plan are chosen by CC, implemented, git tagged, and committed (but not pushed).
4) Do acceptance testing before a git push. Each push results in a new build deployed to production.
I have successfully run the build loop uninterrupted up to 20 times to implement plans that point to multiple feature/platform specs.
Ralph loop GH repo: https://t.co/XVrbjUSqrW
Just attended @gauntletai CTO @ashtilawat's awesome talk, "Building a Software Factory". Highly recommend you take a look at his Github repo (in reply below) and fork it to play with this idea. This factory concept can be applied to any digital creative endeavor, not just software. Talk online at: https://t.co/pzDpLXDKtJ
New image gen model Uni-1 from @LumaLabsAI just dropped and it's architecturally interesting โ autoregressive transformer that reasons through prompts BEFORE generating pixels. Beats Nano Banana on preference tests at 30% lower cost.
Try it: https://t.co/UzmxFYFAC4 Tech details: https://t.co/Py7sUa8HdJ VentureBeat coverage: https://t.co/I3Vp5j0D3i
Are you a software engineer interested in becoming an AI Engineer? Please read on. Do you know engineers who might like to become AI engineers? Please share this post. AI engineering skills are a key differentiator for all of us. I've started the "๐๐ ๐๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ ๐๐ฐ๐ฎ๐ฑ๐ฒ๐บ๐" series of X posts with well defined milestones and live discussions to guide us on our journey towards becoming proficient AI engineers.
๐ ๐๐ ๐๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ ๐๐ฐ๐ฎ๐ฑ๐ฒ๐บ๐, ๐๐ผ๐ต๐ผ๐ฟ๐ #๐ญ, ๐ฃ๐ผ๐๐ #๐ฐ
๐ ๐ ๐ถ๐น๐ฒ๐๐๐ผ๐ป๐ฒ ๐ฐ: ๐๐ด๐ฒ๐ป๐๐ถ๐ฐ ๐๐ โ ๐๐ฟ๐ผ๐บ ๐๐ต๐ฎ๐๐ฏ๐ผ๐๐ ๐๐ผ ๐๐๐๐ผ๐ป๐ผ๐บ๐ผ๐๐ ๐๐ด๐ฒ๐ป๐๐
โ ๐ฆ๐๐ฒ๐ฝ ๐ญ) ๐๐ผ๐๐ป๐ฑ๐ฎ๐๐ถ๐ผ๐ป โ ๐ช๐ต๐ฎ๐ ๐๐ฟ๐ฒ ๐๐ด๐ฒ๐ป๐๐ ๐ฎ๐ป๐ฑ ๐ช๐ต๐ ๐๐ผ ๐ง๐ต๐ฒ๐ ๐ ๐ฎ๐๐๐ฒ๐ฟ? Take Dr. Andrew Ng's "Agentic AI" short course on https://t.co/3WFGXhfyZV. Learn agentic design patterns โ reflection, tool use, planning, and multi-agent collaboration.
โ ๐ฆ๐๐ฒ๐ฝ ๐ฎ) ๐๐๐ถ๐น๐ฑ๐ถ๐ป๐ด ๐ฎ๐ป๐ฑ ๐๐ป๐ด๐ถ๐ป๐ฒ๐ฒ๐ฟ๐ถ๐ป๐ด ๐๐ด๐ฒ๐ป๐๐: Read Anthropic's "Building Effective Agents" guide and their follow-up article "Effective Context Engineering for AI Agents". Build AI agents and feed them the right context (prompts, memory, tool outputs, etc.).
โ ๐ฆ๐๐ฒ๐ฝ ๐ฏ) ๐๐๐ฎ๐น๐๐ฎ๐๐ถ๐ป๐ด ๐๐ด๐ฒ๐ป๐๐: Read Anthropic's "Demystifying Evals for AI Agents". Build evaluation frameworks to measure and improve agent reliability.
โ ๐ฆ๐๐ฒ๐ฝ ๐ฐ) ๐ ๐๐น๐๐ถ-๐๐ด๐ฒ๐ป๐ ๐ฆ๐๐๐๐ฒ๐บ๐ ๐ถ๐ป ๐ฃ๐ฟ๐ฎ๐ฐ๐๐ถ๐ฐ๐ฒ: Read Anthropic's case study on building their multi-agent research system. Learn how multiple specialized agents can coordinate on a complex task.
โ ๐ฆ๐๐ฒ๐ฝ ๐ฑ) ๐๐ด๐ฒ๐ป๐ ๐ฃ๐ฟ๐ผ๐๐ผ๐ฐ๐ผ๐น๐ โ ๐๐ผ๐ป๐ป๐ฒ๐ฐ๐๐ถ๐ป๐ด ๐๐ด๐ฒ๐ป๐๐ ๐๐ผ ๐ง๐ผ๐ผ๐น๐ ๐ฎ๐ป๐ฑ ๐๐ฎ๐ฐ๐ต ๐ข๐๐ต๐ฒ๐ฟ: Read the intro to the Model Context Protocol (MCP) and take https://t.co/3WFGXhfyZV's course on the Agent2Agent (A2A) protocol. MCP connects agents to tools; A2A connects agents to agents.
Please do the above steps and reply below with questions and suggestions.
๐ ๐๐ถ๐๐ฒ ๐๐ ๐ ๐๐ฒ๐๐๐ถ๐ผ๐ป ๐ผ๐ป ๐ช๐ฒ๐ฑ๐ป๐ฒ๐๐ฑ๐ฎ๐, ๐๐ฝ๐ฟ๐ถ๐น ๐ด ๐ฎ๐ ๐ต:๐ฏ๐ฌ ๐ฎ๐บ ๐ฃ๐ฆ๐ง
Please reply below to get the Zoom link to the live AMA session. Bring your questions, discuss progress and connect with peers.
๐ ๐๐ถ๐ป๐ธ๐ (๐ณ๐ผ๐ฟ ๐๐๐ฒ๐ฝ๐ ๐ฎ๐ฏ๐ผ๐๐ฒ):
๐ฆ๐๐ฒ๐ฝ ๐ญ ๐๐ถ๐ป๐ธ: https://t.co/6H7WXbCZlA
๐ฆ๐๐ฒ๐ฝ ๐ฎ ๐๐ถ๐ป๐ธ๐: 1) https://t.co/6HVwmunRsh, 2) https://t.co/t1rOr86U6Z
๐ฆ๐๐ฒ๐ฝ ๐ฏ ๐๐ถ๐ป๐ธ: https://t.co/lLApjoEp0x
๐ฆ๐๐ฒ๐ฝ ๐ฐ ๐๐ถ๐ป๐ธ: https://t.co/fqWNrMHkSN
๐ฆ๐๐ฒ๐ฝ ๐ฑ ๐๐ถ๐ป๐ธ๐: 1) https://t.co/zHQVM09QBj, 2) https://t.co/QPUTw0ggDy
Distributed pre-training of LLMs is getting better. Covenant is a 72B parameter model trained on 1.1T tokens via trustless peers on the internet. It scored 67.4 on the MMLU benchmark beating LLaMa-2-70B (released 3 years ago) which got 63.1 and was trained on ~2T tokens. https://t.co/FuSNaHSiqJ
Last week my goal for 2026 was to build a revenue positive self operating business run by up to 10 autonomous agents. Earlier today I saw Kelly AI: an agent managing an idea factory, a software factory for iOS apps, and a marketing factory generating revenue from day 1. Looks like I need to aim higher! ๐
One thing Iโm thinking about is how much we create vs consume using human or artificial intelligence. I think all knowledge workers can measure and increase their creation-consumption ratio (CCR) by leveraging AI to create software, text, music, audio, images and video. Creation is active whereas most consumption is passive so a higher CCR directly correlates to higher human agency and more bias to action. Optimizing CCR is critical in the AI age.
All knowledge workers can 10x amplify what they love to do *and* 10x reduce what they donโt like to do by leveraging AI. If you are a knowledge worker and not already doing this, figure out a path to do this in 2026. All the tools and technology to enable this is getting better by the day.
Reading Accelerando, written in 2005, feels surreal. The author (Charles Stross) got so many things right, it feels like he time-travelled into the future before writing the book. If you want to prepare for the Singularity, read it: https://t.co/eZwYHrexuX
This is a really interesting development: a software factory: https://t.co/b04YjYyjk7. @strongdm says:
- Code must not be written by humans
- Code must not be reviewed by humans
- If you haven't spent at least $1,000 on tokens today per human engineer, your software factory has room for improvement
Iโm not sure if Iโm ready for $1000/eng/day spend but Iโm intrigued by the factory idea where no code is written or read by humans.
The article links to @strongdm principles, techniques, and products that relate to how they built their software factory.
Definitely worth exploring further.