Wednesday, July 8, 2026
The Shift from Developer Experience to Reliable Agent Intelligence
July 8 · 17 videos
The focus is shifting from DX to AX.
Modal scales to 100,000 sandboxes for reinforcement learning.
Anthropic discovered a reasoning whiteboard inside Claude.
Reliability remains the primary bottleneck for autonomous agents.
Single REPL architectures are replacing complex multi-agent loops.
The UK Ministry of Justice is seeking ground truth on the ground.
“If your agent is making many sequential tool calls or even parallel tool calls, then you've kind of invented a bad scripting language.”
The 100,000 Sandbox Problem: Akshat Bubna, Modal CTO
Akshat Bubna · Latent Space · 59 min
Watch on YouTube →Akshat Bubna explains Modal's evolution into an AI Supercloud. He details how infrastructure must shift from Developer Experience to Agent Experience.
- Modal aggregates capacity across 17 different cloud providers for reliability.
- AI workloads require sub-second cold starts and massive ephemeral scaling.
- The DeFlash speculator provides 2 to 4x inference speedups.
- Observability becomes more critical than code when agents are the primary authors.
- Reinforcement Learning rollouts can demand up to 100,000 concurrent sandboxes.
What do we build now?: Theo Browne, @t3dotgg
Theo Browne · AI Engineer · 16 min
Watch on YouTube →Theo Browne argues that AI models like Mythos have collapsed traditional software tiers. He introduces the Spectrum Strategy for building extensible platforms.
- Projects that once required a full startup can now be a Markdown file on a cron job.
- Engineers should reject the sunk cost mindset and be willing to delete code.
- Small teams can challenge giants by building wide, extensible architectures.
- Mythos represents a jump to orchestration where the model understands itself.
- Professional identity should not be defined by specific languages or tools.
Teaching Computers to Smell: Alex Wiltschko
Alex Wiltschko · The TWIML AI Podcast with Sam Charrington · 59 min
Watch on YouTube →Alex Wiltschko discusses Osmo's work in digitizing the sense of smell. He explains how Graph Neural Networks map molecular structures to scent perception.
- Osmo uses a Principal Odor Map to predict how molecules will smell before synthesis.
- The company has digitized 5.43 million sniffs to train its foundation models.
- Scent digitization could lead to medical diagnostics for cancer or malaria.
- Vertical integration is necessary to build infrastructure for a new modality.
- A Copernican view of intelligence includes chemical communication used by 99 percent of species.
Claude is definitely not conscious
Fireship · Fireship · 4 min
Watch on YouTube →Fireship explores Anthropic's discovery of the J-Space reasoning whiteboard in Claude. This research suggests human-like cognitive structures emerge spontaneously in Transformers.
- The J-Space allows Claude to manipulate thoughts before generating output.
- Researchers used a Jacobian Lens to surgically alter internal reasoning concepts.
- Deleting the J-Space left the model fluent but incapable of logic.
- This discovery mirrors the Global Workspace Theory of human consciousness.
- Mechanistic interpretability tools are moving toward real-time internal modification.
The Pragmatic Engineer AMA
Gergely Orosz · The Pragmatic Engineer · 79 min
Watch on YouTube →Gergely Orosz discusses the shift toward AI-native development and prestige hiring. He shares insights from growing his newsletter to 10,000 paying subscribers.
- Computer Science degrees are becoming critical filters for top-tier companies and visas.
- The pendulum is swinging back toward technical managers who still write code.
- Future-proof careers involve working with AI infrastructure like RAG and fine-tuning.
- The Pragmatic Engineer newsletter reached 1,000 paying subscribers in six weeks.
- Tier 3 companies use LeetCode to test for grit and corporate process tolerance.
Think You Can Build a Game with AI? Think Again!: Danielle An and David Hoe, Meta
Danielle An · AI Engineer · 18 min
Watch on YouTube →Danielle An and David Hoe from Meta discuss the future of AI in gaming. They argue that taste and runtime LLMs will replace simple vibe coding.
- Runtime LLMs allow for non-deterministic NPCs with distinct personalities.
- AI reduces iteration cycles from months to hours through parallel development.
- Creativity is no longer gated by specialized skills like 3D modeling.
- Content safety and automated moderation are essential for AI-generated games.
- The industry is currently at Day Zero of its AI transformation.
AI Found the Trades We Missed
Josh Kale · Limitless Podcast · 30 min
Watch on YouTube →The Limitless Podcast reviews 200 episodes of AI market shifts and hardware returns. They predict recursive self-improvement will manifest by early 2027.
- SanDisk stock saw a 3,500 percent return following a 2025 compute thesis.
- The top three memory manufacturers averaged a 153 percent return in one year.
- Recursive self-improvement is expected to fundamentally change productive work by 2027.
- Economic rents are owned by those who control watts, chips, and cooling.
- The internal Limitless Portfolio returned 4x by investing in podcast guests.
Solving the Blank Canvas Problem: Gusto's AI Co-Founder
Eddie Kim · Y Combinator · 32 min
Watch on YouTube →Eddie Kim explains how Gusto built an agentic co-founder to automate business processes. The product uses Gusto's system of record to perform proactive work.
- Gusto Co-founder automates payroll and compliance via SMS and Slack.
- A five-person team delivered the final product in just ten weeks.
- The initial prototype was built in five hours at an airport.
- Proactive agents find value like tax credits without waiting for prompts.
- Systems of record have a massive advantage in providing necessary AI context.
How to (Actually) Shift Your Identity (to Succeed)
Rob Dial · The Mindset Mentor Podcast · 15 min
Watch on YouTube →Rob Dial explains how identity acts as a thermostat for life outcomes. He provides a framework for lasting self-image transformation.
- 90 percent of people who lose significant weight gain it back within five years.
- Lasting identity change is a long-term process requiring three to four years.
- Confidence is the result of showing up regardless of how one feels.
- Use the Becoming framework to bridge the gap to a new identity.
- A business will never outgrow the internal identity of its owner.
Your agent is blindfolded: Johan Lajili, Poolside AI
Johan Lajili · AI Engineer · 9 min
Watch on YouTube →Johan Lajili discusses the blindfold problem in AI coding agents. He introduces AIX Engineering as the practice of making environments legible for agents.
- Agents fail in legacy environments because they lack sensory feedback loops.
- Spoolside is a CLI tool that gives agents eyes through screenshots and snapshots.
- Engineers must shift from product creation to creating AI-friendly environments.
- Investing in codebase legibility allows agents to run autonomously overnight.
- The market is splitting between those who build feedback loops and those who get garbage.
Building an ACP-Compatible Agent Live — Bennet Fenner, Zed
Bennet Fenner · AI Engineer · 18 min
Watch on YouTube →Bennet Fenner demonstrates building an ACP-compatible agent for the Zed editor. This protocol standardizes how agents interact with development environments, similar to what LSP did for IDEs.
- The Agent Control Protocol (ACP) provides a language-agnostic interface between agents and editors.
- Standardizing tool calling and environment access reduces the integration tax for new AI startups.
- A robust handshake and capability negotiation phase are critical for reliable agent performance.
- ACP enables agents to be portable across different development environments like the Zed editor.
- The protocol shifts AI engineering from monolithic chat features to a modular, decoupled ecosystem.
Teaching Coding Agents to do Spreadsheets - Nuno Campos, Witan Labs
Nuno Campos · AI Engineer · 19 min
Watch on YouTube →Nuno Campos shares how Witan Labs improved spreadsheet agent accuracy to 92 percent. The key was moving from rigid tool calls to a persistent Node.js REPL.
- Spreadsheet accuracy improved from 50% to 92% by moving to a single Node.js REPL with persistent state.
- LLMs struggle with spreadsheets because standard text representations lose visual and structural context.
- Reducing tool count from 15 to 1 eliminated overhead and prevented task timeouts.
- Reliability in financial tasks requires deterministic black-box testing against golden spreadsheets.
- Agent reasoning failures are often actually infrastructure bugs or errors in prompt examples.
Your coding agent doesn't always follow your rules — Talha Sheikh, Checkout.com
Talha Sheikh · AI Engineer · 10 min
Watch on YouTube →Talha Sheikh argues that enterprise AI requires deterministic verification layers rather than just better prompts. He introduces the Vector Harness for real-time enforcement of agent behavior.
- Capability does not equal reliability; smarter models still require feedback loops to follow instructions.
- The Vector Harness uses Claude hooks to trigger real-time deterministic checks on agent output.
- Investing in verification harnesses allows for the use of smaller, cheaper models like Claude Haiku.
- The role of the developer is shifting from writing code to designing verification systems.
- Reliability is the primary barrier to achieving a high ROI for autonomous AI agents.
Running a Chess YouTube Channel entirely by AI — Stephan Steinfurt, TNG
Stephan Steinfurt · AI Engineer · 16 min
Watch on YouTube →Stephan Steinfurt presents an automated pipeline for generating chess commentary videos. The system uses Gemini 1.5 Pro and specialized chess tools to create high-quality content at scale.
- The system generates chess commentary videos for as little as 20-30 cents per video.
- Gemini 1.5 Pro is paired with legal move validators to prevent the model from hallucinating illegal moves.
- The Checks, Captures, and Threats (CCT) framework is used to guide the model's strategic focus.
- The automated channel reached 500,000 views and 4,000 subscribers in just a few months.
- High-quality analysis of user-generated data is a more valuable market than generic automated content.
Shipping Production AI Inside Government — William Tarr, Ministry of Justice
William Tarr · AI Engineer · 18 min
Watch on YouTube →William Tarr and Louis Hogarth explain how the UK Ministry of Justice uses forward-deployed engineering to build AI tools for prisons. It matters because it shows how to bypass bureaucratic resistance by finding ground truth in the field.
- The Justice AI Unit operates as a high-velocity startup within a 95,000-person government organization.
- Engineers spend 2-3 days a week on-site in prisons to observe workflows that policy documents often misrepresent.
- The team uses a forward-deployed model to move from MVP to national rollout in months rather than years.
- Building tools for Victorian-era prisons requires solving for physical constraints like 6-foot walls and no WiFi.
- Presenting real data from working pilots is more effective at neutralizing bureaucratic resistance than using slides.
I Run a Fleet of AI Agents Across Three Machines. Here's What Broke. - Kyle Jaejun Lee, KRAFTON
Kyle Jaejun Lee · AI Engineer · 9 min
Watch on YouTube →Kyle Jaejun Lee discusses the infrastructure failures of running multiple AI agents. He proposes an organizational hierarchy to manage agent state and human cognitive load.
- Managing more than six live agent contexts simultaneously causes human cognitive breakdown.
- Agent infrastructure should mirror an organizational hierarchy with roles like CEO, VP, and worker.
- Moving agent state out of the context window and into disk-based workspaces prevents data loss.
- A reset-not-compact strategy ensures persistence by having agents read mission history from files.
- Future agent platforms must treat compute and tools as an abstraction layer similar to Kubernetes.
Your LLM Deception Monitor Is Broken. The Fix Is in the Training Data - Sachin Kumar, LexisNexis
Sachin Kumar · AI Engineer · 13 min
Watch on YouTube →Sachin Kumar introduces diff SAE to detect sleeper agents in LLMs. This is a critical security advancement for detecting backdoors that standard behavioral testing misses.
- Sleeper agents are backdoors that remain dormant during standard evaluations but trigger malicious behavior later.
- Conventional behavioral monitoring and crosscoders often fail to detect these hidden signals in large models.
- The diff SAE approach isolates directional shifts in model activations during fine-tuning to find backdoors.
- Experiments showed a 40x stronger detection signal compared to existing interpretability methods.
- Integrating a delta monitor into CI/CD pipelines provides a low-cost gate against poisoned models.
References
PeopleAkshat Bubna (https://x.com/akshat_b) · Theo Browne (@t3dotgg) · Alex Wiltschko · Gergely Orosz · Eddie Kim · Rob Dial · Johan Lajili · Josh Kale (@JoshKale) · Ejaaz (@cryptopunk7213) · Isaiah Taylor · Alex Atala · Travis Kalanick · Harj Taggar · James Clear · Volodymyr Giginiak · Mark Zuckerberg · Lenny Rachitsky · Jonas Tyroller · Joel Mainland · Noam Sobel · Terry Tao · Eric Bernhardsson · John Ousterhout · Bernard Baars · Joe Rogan · Danielle An · David Hoe · Bennet Fenner · Nuno Campos · Talha Sheikh · Stephan Steinfurt · Wilhelm Weber · Swyx · Gotham Chess · William Tarr (ai.justice.gov.uk) · Louis Hogarth · Dan James · Owen · David · Kyle Jaejun Lee (x.com/kyleleee_119) · Sachin Kumar
ToolsModal · Mythos · Osmo · Claude · Gusto Co-founder · Spoolside · DeFlash · I6PN · Principal Odor Map (POM) · MeshU · Traycer · Claude Code · Gemini · SanDisk · Valor Atomics · OpenRouter · MindBody · ACP · Zed · Node.js REPL · Vector Harness · Gemini 1.5 Pro · 11 Labs · Lichess · SSH · Git · tmux · Diff SAE · Crosscoders
PapersA Global Workspace in Language Models