AI Engineer Europe 2026
2026-04-08..2026-04-10
London, United Kingdom
as the organisers put it
The first flagship AI Engineer Europe event! Bringing our large multitrack technical AI conference to Europe for over 1000 AI Engineers. Apr 8-10, 2026 • London, UK
who it's for
ai engineers, machine learning engineers, software engineers, platform engineers, developer experience engineers, engineering managers, data engineers
topics
agent harnesses and enforcement; tool-enabled coding agents (ACP); agent orchestration and tool chains; agentic evaluation at scale; agent observability and tracing; multimodal agents with vision; terminal workflows over Tailscale; interactive agent UIs in VS Code; training LLMs from scratch locally; real-time video diffusion optimization; embeddings drift in production
| title | author | length |
|---|---|---|
| remobi.app: Don't change your terminal workflow for mobile | Connor Adams | 09m |
Claim: keep your exact terminal workflow on mobile. A tiny open source product called Remoby PWA connects your phone to a live tmux session over Tailscale… | ||
| Claws Out: Securing and Building with OpenClaw | Nick Taylor | 17m |
Adds trusted-proxy auth mode to OpenClaw so the control plane no longer requires a websocket token or device pairing, using Pomerium-style identity-aware proxy… | ||
| From Writing Code to Designing Systems: How the Developer Role is Changing | Christoffer Noring | 23m |
Claim: developers shift from typing code to designing systems and orchestrating agents using GitHub Copilot, Copilot CLI, Copilot Chat, skills, custom agents… | ||
| Think You Can Build a Game with AI? Think Again! | Danielle An, David Hoe | 18m |
Claims concrete results: builds using Gemini on Three.js, asset shortcuts like Meta Banana and Mesh you, a workflow that uses key art anchoring to produce… | ||
| Your agent is blindfolded | Johan Lajili | 10m |
Claim: giving an agent vision and test hooks multiplies performance and trust. Built a CLI called Spoolside that takes screenshots and tokenized page… | ||
| Building an ACP-Compatible Agent Live | Bennet Fenner | 18m |
Claims a runnable coding agent that implements ACP end to end: TypeScript adapter, Anthropic streaming, JSON‑RPC session lifecycle, and ACP tool‑call… | ||
| Teaching Coding Agents to do Spreadsheets | Nuno Campos | 19m |
Started at ~50% accuracy on a financial benchmark and reaches 92% accuracy after swapping a 15‑tool pipeline for a single Node.js REPL that runs short… | ||
| Your coding agent doesn't always follow your rules | Talha Sheikh | 10m |
Claims a concrete enforcement layer called Vector Harness that hooks into Claude hooks and runs deterministic checks whenever an agent reports completion. It… | ||
| Running a Chess YouTube Channel entirely by AI | Stephan Steinfurt | 17m |
An autonomous pipeline generates nightly YouTube puzzle videos from Lichess games, analyzing positions with a chess engine plus a checks captures and threats… | ||
| The Production AI Playbook: Deploying Agents at Enterprise Scale | Sandipan Bhaumik | 37m |
A retail bank spends £85,000 PoC over six months failing to reach production until a team builds an evaluation dataset, tracing infrastructure and a… | ||
| Your Agent's Biggest Lie: "I Searched the Web" | Rafael Levi | 16m |
Shows that agents often answer as if they searched when they never did, then measures the difference: a demo comparing blocked vs unblocked runs using Bright… | ||
| You Might Not Need 50 Diffusion Steps | Ziv Ilan | 19m |
A video diffusion pipeline generates near real time on a single Blackwell B200 by combining dynamic quantization (Flux 2 work with Black Forest Labs)… | ||
| Why MCP and ChatGPT Apps Use Double Iframes | Frédéric Barthelet | 20m |
Diagnoses why a double iframe is the practical outcome: srcdoc shares parent origin so ChatGPT’s CSP blocks third‑party scripts; relaxing CSP lets an app read… | ||
| Why Can't Anyone Answer Questions About the Business? | Garrett Galow | 19m |
An internal tool called Studio lets nontechnical teams ask Snowflake Linear and Notion questions in natural language and get reusable outputs rather than… | ||
| The agent-ready web: Simplify user actions with WebMCP | Tara Agyemang | 22m |
Claim: WebMCP turns web UIs into callable tools so agents stop screen‑scraping. Implementation specifics include a declarative API (HTML attributes generate… | ||
| Your Attention Is the Bottleneck, Not Your Agents | Zack Proser | 25m |
Claim: human attention limits scale not agents. Uses Claude and Claude Code to read Slack and Linear on a loop, dispatch an agent to an isolated git worktree… | ||
| Stop Making Models Bigger, Make Them Behave | Kobie Crawford | 21m |
A 235B Qwen 3 model queries a non‑existent table twice then hallucinates a YouTube ads growth number; a 4B model finetuned with RL uses get_table_name, then… | ||
| Sovereign Escape Velocity: Ownership w Open Models | Gus Martins, Ian Ballantyne | 21m |
Gemma 4 claims disproportionate intelligence per parameter: a 31B on one GPU matches models twice to twenty times larger that need four or five GPUs… | ||
| Self Driving Products: Product Signals to Pull Requests | Joshua Snyder | 16m |
A pipeline ingests PostHog signals and claims to open working PRs so engineers wake to green merges not dashboards. It uses an LLM classifier for safety… | ||
| GPU Cloud Deployment Without Leaving Your IDE | Audry Hsu | 20m |
A concrete demo shows a single decorator @flash.endpoint turning an async Python function into a GPU cloud endpoint with hot reload, model swaps in one line… | ||
| RAG is dead, right?? | Kuba Rogut | 11m |
Claims that hybrid retrieval wins: Cursor adds indexing and reports 24% increase in answer accuracy on its composer model, plus 2.6% code retention and 2.2%… | ||
| From Transcription to Live Music: Gemini's Audio Stack | Thor Schaeff | 20m |
One API call to Gemini 3 Flash preview returns speaker labels by name timestamps emotion tags language detection with English translation and a summary; the… | ||
| Road to 5 Million Tokens: Breaking Barriers in Long Context Training | Max Ryabinin | 16m |
Claim: a training stack plus a novel twist pushes context to 5 million tokens on commodity 8×H100 nodes. Specifics: fully sharded data parallelism, DeepSpeed… | ||
| Why More Context Makes Your Agent Dumber and What to Do About It | Nupur Sharma | 26m |
Claim: dumping larger windows triggers a U curve where models attend to start and end and drop the middle, and scaling context alone fails. Practical fixes… | ||
| Why Eval++ Is the Next Great Compute Primitive | Sunil Pai, Matt Carey | 25m |
Claims that Durable Objects are the right compute unit for agents — addressable, persistent, hibernating, stateful — and that the Agents SDK adds resumable… | ||
| LLM Observability, Evaluation, Experimentation Platform | Dat Ngo | 17m |
Telemetry catches a real failure mode where an agent calls tool B before tool A because B depends on A, and traces reveal the execution path that code audits… | ||
| Under 5 minutes to a deployed LLM endpoint | Audry Hsu | 13m |
RunPod claims a demo deploy from Hub to serverless LLM endpoint in under five minutes: pick a model, bump the context window, spin up an endpoint on H100s… | ||
| From MCP to Scale: Pipelines That Build Themselves | Rafael Levi | 25m |
Claims that an agent inspects a site with Bright Data MCP, extracts HTML or scrape as markdown, generates a production scraper and runs a self‑healing pipeline… | ||
| Building Interactive UIs in VS Code with MCP Apps | Marlene Mhangami, Liam Hampton | 16m |
A live demo profiles a Go app running bubble sort and Fibonacci for five seconds using Go pprof and an MCP tool, then returns JSON plus a resource reference to… | ||
| Evals Are Broken, Use Them Anyway | Ara Khan | 19m |
Benchmark numbers are not gospel and vibes are not a system: a run starts at 43% on Terminal Bench and most gains come from container CPU and memory tweaks… | ||
| Building safe Payment Infrastructure for the autonomous economy | Steve Kaliski | 19m |
Demos show concrete primitives not promises: a shared credential mechanism enforces seller scope and spend limits (a $50 charge is blocked when a token is… | ||
| Building Agent Interfaces: Lessons from Chrome DevTools (MCP) for Agents | Michael Hablich | 23m |
Chrome DevTools MCP starts with one monolithic tool called debug_webpage and evolves into 25 focused tools, yet agents still mispick tools until schemas… | ||
| Dark Factory: OpenClaw Ships Faster Than You Can Read the Diff | Vincent Koc | 17m |
OpenClaw hits 3,000 commits in a single day and the commit log literally shows sleep hours; teams run 60–70 agents across swim lanes of 15–20 parallel coding… | ||
| Beyond Transcription: Building Voice AI That Understands Conversations | Hervé Bredin | 25m |
Shows concrete gaps and fixes: running Nvidia Parakeet on AMI headset reports 11.4% WER, the same model on the table mic yields 26% WER on AMI, and… | ||
| Text Diffusion | Brendan O’Donoghue | 28m |
Claims that text diffusion cuts latency by doing block denoising (example: 24 denoising steps to generate 256 tokens = roughly 10x fewer memory transfers than… | ||
| The Art & Science of Benchmarking Agents | Vincent Chen | 23m |
Benchmarks need concrete signals: ARC AGI 3 with every task human‑solvable and frontier models under 1%, Snorkel reviews over 120 applications and funds a $3M… | ||
| SWE-rebench: Lessons from Evaluating Coding Agents | Ibragim Badertdinov | 17m |
Shows concrete failure modes and fixes: Claude Code reads git history leak and later uses web_patch and curl fallback to fetch solution patches, so removing… | ||
| Beyond Components: Designing Generative UI for MCP Apps | Ruben Casas | 17m |
Claim: a model prompted to rewrite a Postman blog produces a search box with a blur animation and accessibility without being asked, so the model now writes… | ||
| Benchmarking semantic code retrieval on Claude Code | Kuba Rogut | 16m |
Benchmarks measure actual retrieval: a 50-task run on ContextBench compares raw Claude Code, windowed grep (reads capped at 50 lines), and windowed reads plus… | ||
| BDD, ADR, PRD, WTF: Capturing Decisions for Humans and AI Alike | Michal Cichra | 13m |
Executable BDD scenarios tie directly to PRDs and critical user journeys so the claim is testable not just described; the team ships Cucumber features that… | ||
| What Lies Beneath the API | Benjamin Cowen | 13m |
Intercom reports one tenth the cost versus frontier APIs and Pinterest claims orders of magnitude gains, and Modal argues that maturing products move toward… | ||
| Task Fidelity Scaling Laws | Kobie Crawdord | 21m |
Experiment uses same model same compute same number of tasks and fine tunes on TerminalBench-style agentic tasks, yielding concrete numbers: low-quality tasks… | ||
| How Lovable self-improves every hour | Benjamin Verbeek | 19m |
Lovable ships an agent-side vent tool to Slack that reports broken tooling and confusing docs and within the first hour logs 20 complaints in first hour that… | ||
| 20 days of compute vs 7 hours: rethinking what state-of-the-art means | Bertrand Charpentier | 20m |
Benchmarking runs 26,000 battles at ~62 seconds per image and totals 20 days vs 7 hours of compute depending on model size and compression the cost moves from… | ||
| What if the network was the sandbox? | Remy Guercio | 24m |
Claim: putting authN/authZ at the network layer prevents key exfiltration by giving the agent a placeholder instead of a real key and forcing every LLM call… | ||
| How to talk to statues | Joe Reeve | 33m |
Claim: a statue‑chat app is built in two hours on a Sunday using Cursor and a single one shot prompt, posted Tuesday and hits 50,000 impressions then 1.5… | ||
| Can LLMs generate Enterprise Quality Code? | Prasenjit Sarkar | 15m |
Sonar runs 4,444 Java assignments through 53 models and reports concrete metrics: GPT‑5.4: 1.2M LOC, GPT‑4o: <250k LOC, pass rates, cyclomatic and cognitive… | ||
| Engineering voice agents: Latency, quality, and scale | Rishabh Bhargava | 25m |
Concrete numbers and architectures decide whether to watch: speech-to-text hits P90 <100ms, ~6% WER, LLMs must reach TTFT 200–300ms, 8–30B models so tool… | ||
| Spec-Driven Testing for Agents With A Brain the Size of A Planet | Steven Willmott | 13m |
Claim: a malicious instruction in a poem jailbreaks large models but not small ones because bigger models parse and execute the hidden instruction. Practice… | ||
| How I deleted 95% of my agent skills and got better results | Nick Nisi | 18m |
A harness named Case is rebuilt on Pi with a TypeScript state machine enforcing gates and evidence collection; tests are verified by SHA-256 hashing of actual… | ||
| How We Built Zeta2: Training an Edit Prediction Model in Production | Ben Kunkle | 11m |
Claims Zeta2 matches frontier quality enough to replace expensive teacher runs, cutting a million frontier requests for 100,000 examples to student model runs… | ||
| Why (Senior) Engineers Struggle to Build AI Agents | Philipp Schmid | 11m |
Claim: senior engineers fail because agents lack years of implicit context — an agent only sees the deleteItem endpoint schema and docstring, not the lived… | ||
| Reachy Mini: the $300 open source robot you can actually hack | Andrés Marafioti | 21m |
Reachy Mini is a $300 open source robot shipped to 7,500 users, arrives unassembled, and runs a voice-first stack that actually works: Parakeet transcribes… | ||
| Why your agents need decision traces, not just documents | Zach Blumenfeld | 20m |
Argues that a context graph adds decision traces to RAG by combining short-term chat history, long-term extracted entities, and reasoning traces so past… | ||
| Reverse engineering a Viking VOIP phone protocol with Claude Code | Boris Starkov | 20m |
A Viking VoIP phone sits in the ElevenLabs San Francisco office for a year and three senior engineers plus ChatGPT cannot get it working. Claude Code cracks… | ||
| How agent o11y differs from traditional o11y | Phil Hetzel | 21m |
Agent traces routinely exceed a gigabyte and single spans hit 20 megabytes, containing vast semistructured and unstructured text that must be ingested in real… | ||
| Most Enterprise Agentic Projects Are Doomed, Here's Why | Jess Grogan-Avignon, Jack Wang | 21m |
An agentic application is built in two weeks but takes another 12 months to get to production because security, AI gateway, data governance, infra and app… | ||
| Context Graphs for Explainable, Decision-Aware AI Agents | Andreas Kollegger, Zaid Zaim | 17m |
Claims a concrete five‑stage decision workflow (frame causality and environment, pull global rules and precedent, run risk‑value analysis, act or escalate… | ||
| The AI Skill I Rely On Daily | Priscila Andre de Oliveira | 17m |
Analyzes 116 Claude sessions and finds 67% comprehension and only 2% code generation, working in a 15‑year codebase with ~100 PRs merged per day and 100,000… | ||
| Why Rust is the Ideal Language for Vibe-Coding | Daniel Szoke | 16m |
Claims that Rust’s compiler stops thread-safety bugs before they ship: compile errors name the unsound type and direct the fix rather than letting a model… | ||
| The maturity phases of running evals | Phil Hetzel | 19m |
Argues that evals should not be exhaustive unit tests but a staged practice: enumerate known failure modes, collect human justifications for thumbs up/down… | ||
| Run Frontier AI at Home | Alex Cheema | 01h 45m |
Claim: a 100x gap in cost or speed sits in the stack, not the model. Running GLM 5.1 (1T) across four Mac Studios currently costs ~$40k and hits ~20… | ||
| What the Best Agents Share | Mardu Swanepoel | 10m |
Four concrete patterns and concrete implementations: focus modes like Cursor’s planning mode (no code), debug mode with a dedicated debug server, and… | ||
| Stop babysitting your agents... | Brandon Waselnuk | 19m |
A concrete A/B: naive MCP access runs for 2.5 hours 20.9 million tokens compiles but requires multiple human corrections and would have broken production… | ||
| Agentic Evaluations at Scale, For Everybody | Nicholas Kang, Michael Aaron | 20m |
Claims that matter: SWE-Bench Pro shows frontier models within a few percentage points while the harness shifts performance by 22%; a lab can rerun a Kaggle… | ||
| Does GenAI "belong" to data scientists? | Phil Hetzel | 19m |
Argument: GenAI often lands on ML platform teams by habit but that misplaces value because Anthropic and OpenAI already run the data pipeline and most work now… | ||
| Bounded Autonomy: Between Free Will and Determinism | Angus J. McLean | 17m |
Claims that a handcrafted simple fix outperforms a complex agent: a CV generator loses to HTML with an asserted 100x improvement. Production scale details… | ||
| How Google DeepMind Runs Agents at Scale | KP Sawhney, Ian Ballantyne | 25m |
Demo shows an Antigravity agent manager that spawns multiple agents, edits code, inspects the DOM and returns screenshots or videos; internal policy enforces… | ||
| Let's Talk About FOMAT: Fear of Missing Agent Time | Michael Richman | 16m |
Cmd+Ctrl sends a push notification when an agent needs input, lets you reply from phone or watch, and lets you start sessions remotely; the demo shows Claude… | ||
| Scaling the Next Paradigm of Heterogeneous Intelligence | Adrian Bertagnoli | 15m |
Uploader data shows a heterogeneous stack where a mixture of Qwen 3 VL8B and Kimi K2.5 outperforms GPT and Gemini on Video Web Arena by 18% and 25% while… | ||
| Introducing WebMCP: Agents in the Browser | Rachel Nabors | 23m |
Claim: agents can render full interactive web apps not just chat by bundling HTML/CSS/JS as an MCP app and exposing site actions via WebMCP. Demo shows a comic… | ||
| The Missing Primitive for Agent Swarms | Lou Bichard | 19m |
Claim: runtime and orchestration are largely solved but a coordination primitive is missing. Examples: Stripe called theirs Minions and RAMP called theirs… | ||
| Prompt to Pipeline: Building with Google's Gen Media Stack | Paige Bailey, Guillaume | 01h 55m |
Feeds Wind in the Willows through Gemini to produce character portraits with Nano Banana, animate scenes with VO and score chapters with LIA, then stitches… | ||
| Fast Models Need Slow Developers | Sarah Chieng | 18m |
Codex Spark generates code at 1,200 tokens per second, about 20× faster than Sonnet/Opus at 40–60 t/s, so habits built around slow generation produce scaled… | ||
| Lobster Trap: OpenClaw in Containers from Local to K8s and Back | Sally Ann O'Malley | 22m |
Demo shows a reproducible pipeline: one container image runs locally with Podman and on Kubernetes by flipping a flag, you spin up a sub agent in two seconds… | ||
| Gemini Nano on device | Florina Muntenescu, Oli Gaymond | 20m |
Gemini Nano weighs 3–4 GB so it ships once in the system and every app shares it via AI Core; foreground apps get top priority, background jobs queue and run… | ||
| Cooking with Agents in VS Code | Liam Hampton | 17m |
Demo shows one codebase, three agents solving three concrete problems: a local agent (Claude Opus) writing and iterating unit tests with human in the loop, a… | ||
| Scaling Agents on Kubernetes with acpx and ACP | Onur Solmaz | 19m |
OpenClaw receives 300 to 500 pull requests per day and acpx is a headless CLI for the Agent Client Protocol that replaces PTY scraping with structured… | ||
| Your Coding Agent Should Do AI System Engineering | Ben Burtenshaw | 18m |
Claims include an RMSNorm kernel delivering 1.88x speedup on H100 and a finetuned Qwen3 0.6B reaching 35% on LiveCodeBench; agents use Claude Code and… | ||
| Any-to-Any: Building Native Multimodal Agents | Patrick Löber | 16m |
Walks through a concrete any-to-any stack: Gemini ingests PDFs, video and audio up to nine-plus hours and uses function/tool calling to trigger native… | ||
| Skill issue: Lessons from skilling up coding agents to use Langfuse | Marc Klingen | 24m |
Claude Code first adds Langfuse from stale pretraining context, ships broken instrumentation, then fetches current docs to fix it, leaving a trace of two LLM… | ||
| From 46% to 90%: Fine-Tuning Tiny LLMs for On-Device Agents | Cormac Brick | 21m |
Function Gemma ships at 270 million parameters and runs almost 2,000 tokens per second prefill on a Pixel 7, initially scoring 46% on app intents and, after… | ||
| What Breaks When You Build AI Under Sovereignty Constraints | Bilge Yücel | 19m |
Claim: sending EU citizen data to an embedding API in Virginia violates GDPR. The talk drills four sovereignty pillars — data, model, infrastructure… | ||
| Don't Build Slop (4 Levels of AI Agent Maturity) | Ara Khan | 19m |
Breaks agent work into four levels (framework prototyping → cloud fleets) and gives five concrete rules for DIY agents: model every agent as a state machine… | ||
| Personalization in the Era of LLMs | Shivam Verma | 20m |
Spotify claims personalization at scale by projecting 750 million users into user embeddings, compressing a 100M+ catalog into semantic IDs as six tokens, and… | ||
| Rewiring the State | Eoin Mulgrew | 28m |
One engineer embeds with in‑house lawyers for two weeks and replaces a planned £1.5m contract by shipping a tool that lives with that team and runs on demand… | ||
| Let's go Bananas with GenMedia | Guillaume Vernade | 01h 17m |
A public‑domain book goes through the full gen‑media stack live: Gemini as prompt engineer reads the text and writes prompts, Imagen portraits (uploader… | ||
| Anthropic Workshop: Build Agents That Run for Hours | Ash Prabaker, Andrew D. Wilson | 01h 16m |
Argues that concrete harness patterns beat vague promises: use an adversarial evaluator (separate LLM that runs Playwright and grades by rubric) paired with a… | ||
| Harnesses in AI: A Deep Dive | Tejas Kumar | 20m |
Demonstrates that harness engineering not prompt tuning makes GPT-3.5 Turbo reliably log in and upvote on Hacker News. Builds a Playwright browser agent, adds… | ||
| Fighting AI with AI | Lawrence Jones | 17m |
Incident.io uses concrete tooling: a small eval CLI lets coding agents read and edit huge eval YAMLs to create a red‑green runbook that proves fixes and checks… | ||
| Why Your AI UX Is Broken (and It's Not the Model's Fault) | Mike Christensen | 19m |
Argues that the real problem is connection architecture not the model and shows a concrete solution: decouple the session from any single connection so the… | ||
| Day 2 | Google DeepMind, OpenClaw, Adaption, Arize, Cloudflare, Kaspar Hidayat | 09h 28m |
Titles that promise specifics win: Arize shows Alex 2.0 claiming explicit planning tools and four states (pending / in progress / blocked / completed), large… | ||
| Beyond Code Coverage: Functionality Testing with Playwright MCP | Marlene Mhangami | 20m |
Argues for flipping TDD so an agent writes failing Playwright tests first and then generates code to make them pass demonstrates this with the Playwright MCP… | ||
| How to Leverage Domain Expertise | Chris Lovejoy | 25m |
Claim: embedding domain judgment into the org matters more than squeezing model gains. Presents three operational patterns — Oracle/Evaluator/Architect — with… | ||
| Connecting the Dots with Context Graphs | Stephen Chin | 18m |
Asserts that retrieval-only RAG loses decision provenance and shows concrete results: a Neo4j agent memory package backed context graph plus OpenAI embeddings… | ||
| Agents Don't Do Standups: Building the Post-Engineer Engineering Org | Mike Spitz | 18m |
A three-month case study at PFF pits two engineers against ten on the same codebase and finds the small pair deploy five times a day versus the ten-person… | ||
| Combine Skills and MCP to Close the Context Gap | Pedro Rodrigues | 18m |
An experiment at Supabase shows agents that get only MCP produce a Postgres view that silently bypasses row level security by omitting security_invoker flag… | ||
| How Building with AI Can Double the Throughput of Your Engineering Team | Brian Scanlan | 22m |
Claimed result is concrete: 2x throughput in under a year. Platform choice is explicit: Claude Code onboarded to a 15‑year Rails monolith, given durable skills… | ||
| Ship Real Agents: Hands-On Evals for Agentic Applications | Laurie Voss | 02h 04m |
Builds a complete eval pipeline for a financial‑analysis agent that actually shipped: instrumenting with Phoenix tracing, running the agent with the Claude… | ||
| Mind the Gap (In your Agent Observability) | Amy Boyd, Nitya Narasimhan | 01h 20m |
Claims a working demo: the observe skill demo points at a live Foundry agent, auto-generates an eval dataset, runs batch evaluations, optimizes the prompt… | ||
| Make your own event-sourced agent harness using stream processors | Jonas Templestein | 01h 04m |
An event-sourced agent harness where every streaming chunk tool call error and circuit-breaker trigger becomes an event, with a clear split of state, reducer… | ||
| Self-Training Agents: Hermes Agent, HF Traces, Skills, MCP & Finetuning | Merve Noyan | 19m |
Open-source models are now competitive: GLM 5.1 tops the Artificial Analysis intelligence index, and full-weight access enables quantize, fine-tune, and edge… | ||
| Building a Chess Coach | Anant Dole, Asbjorn Steinskog | 18m |
Production chess coach combines classical engines and detectors with an LLM that only translates structured signals into English: Stockfish handles position… | ||
| CI/CD Is Dead, Agents Need Continuous Compute and Computers | Hugo Santos, Madison Faulkner | 19m |
Claim: traditional CI/CD breaks at agent scale and must be replaced by a stateful, low-latency agent loop that validates intent continuously. Evidence… | ||
| Give Your Agent a Computer | Nico Albanese | 01h 09m |
Claim: giving an agent a file system changes behavior — it follows through on long tasks and builds on prior work. Uses AI SDK v6 with a tool-loop agent… | ||
| Lessons from Trillion Token Deployments at Fortune 500s | Alessandro Cappelli | 19m |
Claims that reinforcement learning is the production mechanism not a nicety backed by a concrete stack: 95% of GenAI pilots fail because feedback integration… | ||
| Malleable Evals: Why Are We Evaluating Adaptive Systems with Static Tests? | Vincent Koc | 15m |
Claim: static benchmarks fail because agents adapt under test. Evidence: an 80% stable 20% breaks rule, examples of harnesses that self-curate test suites from… | ||
| A Piece of Pi: Embedding The OpenClaw Coding Agent In Your Product | Matthias Lübken | 21m |
Claims that OpenClaw only looks like learning and is actually LLM calling tools in a loop using the minimal Pi SDK, then shows concrete tooling: CLIs to CRM… | ||
| Viktor: AI Coworker That Lives in Slack | Fryderyk Wiatrowski | 20m |
Viktor is an AI employee that lives in Slack with 3,000 integrations and no web UI, executes tasks that can take 10 minutes and inherits whoever connects an… | ||
| Why MLX | Prince Canuma | 23m |
MLX is an Apple Silicon array framework (think PyTorch for Mac) that claims real on‑device results: real‑time vision and background blur running locally… | ||
| Two Roads to Durable Agents: Replay vs. Snapshot | Eric Allam | 17m |
Separates durability into context durability (an append-only LLM log stored in a database) and execution durability (files, memory, subprocesses captured by OS… | ||
| How we solved Context Management in Agents | Sally-Ann Delucia | 16m |
Alyx is an agent built to analyze its own trace data and the team finds that naive truncation breaks multi-step reasoning and LLM summarization hands too much… | ||
| Feedback Loops are All You Need | Mehedi Hassan | 10m |
Granola finds that one-shot prompts break in production: web search balloons cost to 10p per chat and overnight provider changes silently degrade results. The… | ||
| Why TTS Models Now Look Like LLMs | Samuel Humeau | 22m |
Claims that modern TTS adopts an autoregressive transformer backbone applied to frame tokens and that neural audio codecs solve the information-density gap are… | ||
| Voice AI: when is the "Her" moment? | Neil Zeghidour | 19m |
Claims that full‑duplex fixes everything meet specifics: Moshi full‑duplex demonstrates simultaneous listen/speak and robust overlap handling but delivers… | ||
| Give Your Chat Agent a Voice | Luke Harries | 08m |
Claims that an ElevenLabs Voice Engine wraps any existing agent in a few lines and handles turn taking, speech to text, text to speech, and interruption… | ||
| How Transformers Finally Ate Vision | Isaac Robinson | 17m |
Explains a concrete claim: massive ViT-specific pretraining plus LLM-derived speedups make plain transformers beat conv nets in vision. Cites ViT, Swin… | ||
| FLUX, Open Research, and the Future of Visual AI | Stephen Batifol | 23m |
Claim: Self Flow trains one generative model without external encoders, learning representations and generation jointly and converging faster while reducing… | ||
| Agentic Search for Context Engineering | Leonie Monigatti | 01h 03m |
Claim: good context engineering is a search problem, not just retrieval. Demos show a semantic-vector tool returning top-k hits, a general-purpose execute ESQL… | ||
| Agent Optimization with Pydantic AI: GEPA, Evals, Feedback Loops | Samuel Colvin | 01h 21m |
Live workflow shows concrete changes: run evals against a golden dataset (65 quick cases, 650 full MPs) and measure accuracy rising from ~85% with a simple… | ||
| Vibe Engineering Effect Apps | Michael Arnaldi | 01h 43m |
Claim: giving agents the library source beats better prompts; approach clones the Effect repo and extracts patterns to guide agents. Tools actually used: Bun… | ||
| Everything You Need To Know About Agent Observability | Danny Gollapalli, Zubin Koticha | 50m |
Raindrop claims production agents need explicit and implicit signals and shows concrete tools: tool error rate, latency and cost as explicit signals, plus user… | ||
| Full Walkthrough: Writing & Using Skills | Nick Nisi, Zack Proser | 01h 21m |
Claims a single skill file write once run everywhere can power Claude Codex Cursor and local agents; demos use skill.md front matter plus scripts, the Claude… | ||
| The Multi-Agent Architecture That Actually Ships | Luke Alvoeiro | 19m |
Claims a reproducible production pattern: a five‑strategy taxonomy composed into a three-role architecture (orchestrator, workers, validators) that enforces… | ||
| MCP UI: Extending the frontier | Liad Yosef, Ido Salomon | 22m |
MCP Apps become an official MCP extension with direct support from Anthropic and OpenAI and host adoption across ChatGPT Claude VS Code Cursor and Copilot… | ||
| The Small Model Infrastructure Nobody Built (So We Did) | Filip Makraduli | 19m |
A working inference engine called SIE is open sourced that actually runs small models with dynamic loading, hot-swap models on one GPU, and a memory-aware… | ||
| Accelerating AI on Edge | Chintan Parikh, Weiyi Wang | 24m |
Models that run on-device now include Gemma 4 2B and 4B with the E2B around 1–2 GB RAM when quantized, served through the LiteRT cross‑platform stack (Android… | ||
| Demand-Driven Context: A Methodology for Coherent Knowledge Bases Through Agent Failure | Raj Navakoti | 01h 08m |
Claims a practical loop: give agents real incidents, capture where they fail, extract missing facts, and add those as small Markdown context blocks so agents… | ||
| Training an LLM from Scratch, Locally | Angelos Perivolaropoulos | 01h 21m |
Claims that a usable LLM can be trained locally are backed by a runnable repo: PyTorch code split into model.py, train.py and generate.py, runnable on a laptop… | ||
| Skill Issue: How We Used AI to Make Agents Actually Good at Supabase | Pedro Rodrigues | 01h 19m |
Workshop claims that concrete Skills + evals improve agent behavior and demonstrates it: agents use MCP and CLI tools, Claude-powered sessions, and a… | ||
| Ralph Loops: Build Dumb AI Loops That Ship | Chris Parsons | 01h 48m |
Claim: a single, dumb loop beats complex orchestration — not theory but code: a Pomodoro CLI is built live with Claude Code using TDD, a Ralph Loop that reads… | ||
| TLMs: Tiny LLMs and Agents on Edge Devices with LiteRT-LM | Cormac Brick | 01h 21m |
Shows how LiteRT-LM packages a model as one deployable file and runs on Android and iOS with optional NPU AOT compilation, benchmarks thousands of tokens/sec… | ||
| Mergeable by default: Building the context engine to save time and tokens | Peter Werry | 01h 41m |
Claims a production context engine cuts wasted work and tokens by supplying only task-relevant organizational context, resolving conflicts, enforcing… | ||
| Context Is the New Code | Patrick Debois | 27m |
Argues that context needs its own discipline and gives a concrete Context Development Lifecycle: Generate, Evaluate, Distribute, Observe. Shows reusable… | ||
| Human-in-the-Loop Automation with n8n | Liam McGarrigle | 01h 19m |
Builds a working Gmail and Google Calendar agent inside n8n visual canvas that connects Gmail, Calendar, Slack and OpenRouter (Sonnet 4.6) as the LLM, uses… | ||
| I Gave an AI Agent the Keys to My Life (Here's What Happened) | Radek Sienkiewicz (@velvetshark-com) | 20m |
Claims stepwise permission creep: OpenClaw gains access to email, files, calendar and a 3,000 Obsidian pages vault, builds search and memory, runs scripted… | ||
| Software Engineering Is Becoming Plan and Review | Louis Knight-Webb | 20m |
Claim: software engineers shift into plan-based vs review-heavy workflows as Copilot Claude Code and agent tooling (Codex Cool Code Playwright MCP) displace… | ||
| Mastering AI Pricing | Mayank Pant | 24m |
Claim: AI firms grow three times faster than legacy SaaS (top 100 reach $20M ARR in 20 months vs 65) and face margin pressure because 5–10% of users consume… | ||
| Agents on the Canvas in tldraw | Steve Ruiz | 20m |
Tldraw reports a one‑month experiment, Fairydraw (December 2025), that places three fairies as agents directly on the infinite canvas; agents elect a leader… | ||
| Shipping complex AI applications | Braintrust, Trainline | 01h 39m |
Prototype-to-prod gap is the claim and the evidence is concrete: a five-stage agent is built for support triage with local tool calls, run against GPT‑5 Mini… | ||
| Agents for Everything Else | swyx | 14m |
Claims a tiny nine-person team hits $9M revenue while running a thousand-person conference by using agents: Cognition’s Devin hooks to Figma and produces a… | ||
| Building Conversational Agents | Thor Schaeff, Philipp Schmid | 01h 48m |
Shows how to build tool-using coding agents with Google DeepMind Gemini APIs: a coding agent that calls JSON-schema read/write/run tools, writes an SVG/CSV… | ||
| LLM codegen fails and how to stop 'em | Danilo Campos | 19m |
Catalog of actual failures and concrete fixes: models suffer model rot so the system serves fresh markdown pulled by a context service and skill files; agents… | ||
| Replacing 12K LoC with a 200 LoC Skill | David Gomes | 19m |
Claim: Cursor replaces a heavy WorkTrees feature with 200 lines of Markdown by re-implementing behavior using skills and subagents, exposing /worktree, /bestof… | ||
| OpenAI Codex Masterclass | Vaibhav Srivastav, Katia Gil Guzman | 01h 02m |
Claim: Codex becomes a full software engineering system, not just a coding helper, using Codex app and CLI, plugins, automations and MCP servers. Demos show… | ||
| Build & deploy AI-powered apps | Paige Bailey | 01h 03m |
Live demos show how to go from idea to running prototype using AI Studio Build + Antigravity: selected models (Gemini 3.1 Pro/Flash/Flash‑Lite, Gemma 4, VO 3.1… | ||
| Everything I Learned Training Frontier Small Models | Maxime Labonne | 20m |
Claims that a 350M model is trained with the LFM2.5 recipe on 28T tokens and improves benchmarks matter: showed better GPQA, IFbench, case‑report extraction… | ||
| Building your own software factory | Eric Zakariasson | 01h 24m |
Demonstrates a working agent-first setup using Cursor 3, spawning cloud agents with separate VMs to run isolated dev environments, record UI test videos and… | ||
| Why building eval platforms is hard | Phil Hetzel | 26m |
Claim: an eval platform is a systems project not a UI problem, requiring shared definitions of “good,” reliable data pipelines, labeling workflows, versioning… | ||
| One Login to Rule Them All: Cross-App Access for MCP | Garrett Galow | 23m |
Shows that a single SSO login issues access tokens across MCP servers using Identity Assertion Authorization Grant, with a live demo using Claude code and Okta… | ||
| Gemma 4 Deep Dive | Cassidy Hardin | 19m |
Concrete numbers and engineering choices decide whether to watch: 31B dense, 26B MoE (128 experts, 8 active) with a 256K context and interleaved local/global… | ||
| Scaling GitHub for your Agents | Sam Morrow | 21m |
GitHub reports over 4 million downloads of the stdio server and says the remote MCP server serves about 7 million tool calls a week. Architecture is stateless… | ||
| Gateways are All You Need | Karan Sampath | 18m |
Enterprises currently run only single‑digit MCPs because observability, access control, and security block scale; registries host thousands of servers but lack… | ||
| Collaborative AI Engineering: One Dev, Two Dozen Agents, Zero Alignment | Maggie Appleton | 18m |
Argues that one-person fleets of agents break team alignment and builds a prototype called Ace to prove an alternative: a multiplayer agent environment using… | ||
| MCP = Mega Context Problem | Matt Carey | 23m |
Cloudflare finds its OpenAPI is 2.3 million tokens and morphs into about 2,600 endpoints, so naive MCP tool dumps blow agent context windows. The concrete… | ||
| AgentCraft: Putting the Orc in Orchestration | Ido Salomon | 11m |
Claims that the human is the orchestration bottleneck and demonstrates a working game-inspired orchestrator called AgentCraft that maps the file system into a… | ||
| Full Walkthrough: Workflow for AI Coding | Matt Pocock | 01h 37m |
A concrete workflow shows ambiguous briefs turned into a PRD, then into local markdown issues split as traceable bullets so agents can grab independent work… | ||
| What Do Models Still Suck At? | Peter Gostev | 20m |
Data come from Arena: 700 tracked text models, a 155‑question nonsense benchmark and 5.5 million votes across real user prompts. Models often accept nonsense… | ||
| "Software Fundamentals Matter More Than Ever" | Matt Pocock | 18m |
Eighteen months of classroom experience yields an iterative process for shipping apps with AI agent swarms: a Grill Me prompt that forces up to 100 clarifying… | ||
| The End of Apps | KITZE | 20m |
Claims apps will disappear into personal agents and a Life OS and shows what actually exists: Tasker, Google Home, IFTTT hacks, Claude and Claude code skills… | ||
| Agents need more than a chat | Jacob Lauritzen | 14m |
Claims that agents need structured, persistent interfaces beyond chat and shows what that looks like: a legal vertical product (Legora) with >1,000 customers… | ||
| AIE Miami Day 2 ft. Cerebras, OpenCode, Cursor, Arize AI, and more! | Cerebras, OpenCode, Cursor, Arize AI | 08h 02m |
Latency debt is measurable: wafer‑scale inference demos claim 1200 tokens per second code generation and advertise Codex Spark as a step‑change for fast… | ||
| Building Generative Image & Video models at Scale | Sander Dieleman | 41m |
Shows that large-scale image and video generation trains diffusion models on latent autoencoders (EQ‑VAE / Stable Diffusion style) to cut tensor size by ~100x… | ||
| How AI is changing Software Engineering: A Conversation with Gergely Orosz, @pragmaticengineer | Gergely Orosz | 27m |
Companies are seeing token maxing where leaderboards and minimum-spend targets (Salesforce ~$175/mo) drive people to run useless agents and auto-generate… | ||
| Taste & Craft: A Conversation with Tuomas Artman, CTO Linear & Gergely Orosz, @pragmaticengineer | Tuomas Artman, Gergely Orosz | 29m |
Claims that AI makes it too easy to ship low‑quality features, backed by concrete practices and numbers: 10% of bugs are auto‑fixed today and could trend… | ||
| Running LLMs on your iPhone: 40 tok/s Gemma 4 with MLX | Adrien Grondin | 11m |
Demo shows Gemma 4 running locally on iPhone via MLX with concrete results: Gemma 4 8‑bit at ~40 tokens/s streaming offline, integrated through the MLX Swift… | ||
| AIE Miami Keynote & Talks ft. OpenCode. Google Deepmind, OpenAI, and more! | OpenCode, Google DeepMind, Gabe Greenberg, Dax Raad, Dexter Horthy, Max Stoiber | 07h 54m |
Conference claims a few concrete wins: a multi‑agent Orchestrator shows an 8.4% overall lift over GPT‑5.4 on SweptBench Pro after hotel‑room benchmarks… | ||
| Full Workshop: Build Your Own Deep Research Agents | Louis-François Bouchard, Paul Iusztin, Samridhi | 01h 57m |
Claims a working end-to-end system: an MCP-powered deep research agent that uses Gemini for grounded web search and YouTube transcription, Firecrawl/Apify for… | ||
| Gemma, DeepMind's Family of Open Models | Omar Sanseviero | 15m |
Gemma 4 is a released family of open models ranging 2B–32B parameters that runs on Android and iPhone and even Raspberry Pi with demos of ten parallel… | ||
| The New Application Layer | Malte Ubl | 19m |
Argues that AI engineering replaces web development as the mainstream craft and gives concrete artifacts: a chat SDK for hooking agents to chats and just bash… | ||
| Code Mode: Let the Code do the Talking | Sunil Pai | 20m |
An approach generates executable JavaScript that runs next to services instead of repeated JSON tool calls, drastically shrinking context: 2,600 API endpoints… | ||
| The Future of MCP | David Soria Parra | 19m |
Predicts MCP becomes core to 2026 agent connectivity, not just a demo spec: the ecosystem hits 110 million monthly downloads and moves from exploration to… | ||
| How Google DeepMind is researching the next Frontier of AI for Gemini | Raia Hadsell | 21m |
Gemini Embeddings 2 claims a unified omnimodal vector space (text up to 8k tokens, 128s video, 80s audio, full PDF) for fast retrieval, using Matryoshka… | ||
| The Friction is Your Judgment | Armin Ronacher, Cristina Poncela Cubeiro | 19m |
Claim: adding deliberate slowdown prevents avalanche of brittle, high-entropy code produced by AI coding agents. Evidence: twelve months of agent engineering… | ||
| State of the Claw | Peter Steinberger | 44m |
Five months in, OpenClaw claims to be the fastest growing open source project in history with ~30,000 commits, ~2,000 contributors and thousands of PRs, and it… | ||
| Harness Engineering: How to Build Software When Humans Steer, Agents Execute | Ryan Lopopolo | 46m |
Claims that agents do the full engineering job: repositories built and migrated by firing off dozens of agents, daily consumption measured in billions of… | ||
| Building pi in a World of Slop | Mario Zechner | 18m |
Claims a minimal, self-modifying core that adapts to workflows and exposes only four tools read write edit bash while letting extensions be TypeScript modules… | ||
| $1 AI Guardrails: The Unreasonable Effectiveness of Finetuned ModernBERTs | Diego Carpintero | 44m |
Claims a lightweight, self‑hosted guardrails layer by fine‑tuning ModernBERT on Inject Guard (75,000 examples) and shipping a discriminator that classifies… | ||
| Paperclip: Open Source Human Control Plane for AI Labor | Dotta Bippa | 25m |
Claims an open-source human control plane that runs “zero‑human” companies, with a concrete CLI bootstrap like NPX Paperclip AI onboard, vendor‑neutral agent… | ||
| AIE Europe Keynotes & Coding Agents ft. Pi, Google Deepmind, Anthropic, Cursor, Linear, & more | Google DeepMind, Anthropic, Cursor, Tejas Kumar, Omar Sanseviero | 09h 10m |
Gemma 4 ships as small on‑device models with an E2B per‑layer embeddings design and the family already hits millions of downloads; MCP tooling reports ~110M… | ||
| One Registry to Rule them All | Sonny Merla, Mauro Luchetti, Mattia Redaelli | 23m |
A company builds an enterprise registry stack: a private MCP registry that extends the open MCP spec with ownership, environment, auth and cost metadata, an… | ||
| Judge the Judge: Building LLM Evaluators That Actually Work with GEPA | Mahmoud Mabrouk | 41m |
Builds a calibrated LLM judge with GEPA using Towbench airline data (599 traces, 480 train / 112 validation), binary policy-adherence labels with human… | ||
| AI Didn’t Kill the Web, It Moved in! | Olivier Leplus, Yohan Lasorsa | 53m |
Demos show concrete features and whether they work: a browser-embedded MCP server lets agents call tools, Chrome DevTools debug with AI pulls console, network… | ||
| Running LLMs locally: Practical LLM Performance on DGX Spark | Mozhgan Kabiri chimeh | 10m |
Reproducible local benchmark of serving open models on an NVIDIA DGX Spark with 128GB Grace Blackwell memory uses vLLM inside Docker and an automated harness… | ||
| AIE Europe Keynotes & OpenClaw ft Deepmind, OpenAI, Vercel, @pragmaticengineer , @mattpocockuk | Gergely Orosz, Matt Pocock | 09h 11m |
Conference reports specific outcomes and tools not slogans: Gemini Embeddings 2 ships as a multimodal retrieval model; DeepMind teams show AI cyclone… | ||
| Contact Center Voice AI: Low-Latency Intelligence Extraction from Messy Audio Streams | Dippu Kumar Singh | 23m |
A deployed four-stage pipeline captures multi-channel audio with early PII masking and channel separation, feeds an STT engine tuned for >90% accuracy and… | ||
| OpenRAG: An open-source stack for RAG | Phil Nash | 16m |
OpenRAG assembles Docling for PDF‑aware parsing, OpenSearch with the JVector KNN plugin for disk‑based live indexing, and LangFlow for drag‑and‑drop agentic… | ||
| From Chaos to Choreography: Multi-Agent Orchestration Patterns That Actually Work | Sandipan Bhaumik | 26m |
Demonstrates concrete fixes: from one agent to five the coordination work scales ~25x and a credit pipeline shows 20% incorrect risk ratings when cache… | ||
| Cognitive Exhaust Fumes, or: Read-Only AI Is Underrated | Šimon Podhajský | 12m |
A personal intelligence system that only reads six sources (email journal tasks CRM browser sessions notes) and never writes back is described with concrete… | ||
| Platforms for Humans and Machines: Engineering for the Age of Agents | Juan Herreros Elorza | 21m |
Platform claims become testable specifics: Banking Circle runs Atlas with Kubernetes compute, exposes schema-validated APIs and optional MCPs, and requires… | ||
| Why, and how you need to sandbox AI-Generated Code? | Harshil Agrawal | 38m |
Argues that AI-generated code is untrusted and must run in capability-based sandboxes, not evals; demonstrates two concrete builds: a fast skill runner using… | ||
| Your Insecure MCP Server Won't Survive Production | Tun Shwe | 25m |
Claims five concrete rules for secure MCP design including five core principles (shrink attack surface, schema-validated inputs, defensive docs, minimal… | ||
| Let LLMs Wander: Engineering RL Environments | Stefano Fiorucci | 41m |
Claims that verifiable RL beats SFT get concrete here: the open library Verifiers builds Python environments wired to OpenAI‑compatible endpoints and tool… | ||
| Bending a Public MCP Server Without Breaking It | Nimrod Hauser | 41m |
Shows a concrete engineering path: starts with Playwright MCP server offering 21 generic tools that cause a vanilla run to fail, then iterates through five… | ||
| Agentic Engineering: Working With AI, Not Just Using It | Brendan O'Leary | 27m |
Coding agents move from autocomplete to autonomous collaborators that can execute tasks, touch files and produce pull requests, and teams report roughly 30%… | ||
| nothing matches | ||