About this show
We turn high signal in-person events for the top AI engineers, founders, leaders, and researchers in the world into the best free learning opportunities for millions around the world here on YouTube. Your subscribes, likes, comments, speaking, attendance, or sponsorships goes a long way toward making our biz model sustainable indefinitely.
We strongly believe this industry deserves a better class of community and that we know how to do this well; we just need your support.
Videos come from the channel's public YouTube feed. The artwork is the newest video thumbnail. Last updated . Names, artwork and text belong to their publishers.
Cosmos
Similar by topic and creator
- AppleInsider
Shares the topic Technology. - AMD
Shares the topic Technology. - 9to5Mac
Shares the topic Technology.
Videos
Your Agent Hates Your CLI - Ayush Bhardwaj, Allos AI
· 12 min
An agent can hide a failed test behind a successful shell pipeline. It can also trim away the error message and confidently report that everything worked. Ayush Bhardwaj examined roughly 20,000 tool calls from his own coding sessions and found a product roadmap in the commands agents wrote immediately afterward. Repeated uses of head, tail, grep and Python showed where tools returned too much information, exposed the wrong fields, or made the agent reconstruct the same query again. The examples come from his own traces and his chat history CLI, which lets agents retrieve context from earlier coding sessions. The fixes are small and concrete: return the messages around a search match, offer compact search results, inspect several sessions in one call, and put a useful failure summary at the end of the output. But implementing a flag does little if the skill file still teaches the old interface. In Ayush's case study, agents kept passing a flag that did nothing until the instructions changed. He also explains how to turn recorded calls into replay tests, reserve examples for evaluation, and check whether better output actually improves answers and reduces turns. Smaller responses alone do not establish a faster or cheaper workflow. The useful feedback is already in the transcript: what the agent tried, what it discarded, and what it had to ask again. Speaker info: Ayush Bhardwaj, Allos AI - https://ayushb.me/ - https://github.com/ay-bh/chat-history Timestamps: 0:00 - The workarounds hiding in agent calls 1:37 - Auditing 20,000 calls and misleading exit codes 2:59 - The chat-history CLI case study 4:32 - Smaller slices, compact results and batch requests 6:33 - Skills and help text are the interface 7:36 - Turn transcripts into replay tests 9:19 - Agent mode and measuring useful outcomes 10:14 - Design the interface, then test against real traces
Not rated yetEvery Correction You Give Your Agent Is Data - Brian Douglas, Paper Compute
· 12 min
A Rust program can compile and still violate the decisions an experienced teammate would catch immediately. Brian Douglas finds that missing knowledge in moments when engineers push back on their coding agents. At Paper Compute, he captures agent sessions, labels corrections and uses those traces to recover the reasoning that otherwise disappears inside a local coding harness. His example starts with a gap in his own Rust expertise: accepting plausible generated code leads to objections later, when a teammate reviews the pull request. Those earlier objections become useful training material only if someone preserves and labels them. Brian demonstrates session capture with Tapes and Paper Console, then runs a Twin Review prototype against a pull request using examples of a teammate's feedback. He discusses labels for pushback, successful sessions and regressions, and shows how the same records can support skills, evaluations or model adaptation. The prototype is a demonstration, not a measured comparison against an independent evaluation set, and Brian corrects himself during the recording to clarify that Twin Review is not publicly released because its data is sensitive. His practical starting point is simpler than training a model: collect the traces you already generate, find the corrections that encode your team's conventions, and turn those into a skill you can test. The durable asset is the connection between a decision, the expert's objection and the resulting outcome. Speaker info: Brian Douglas, Paper Compute - https://b.dougie.dev/ - https://briandouglas.me/ - https://tapes.dev/ Timestamps: 0:00 - Engineering judgment and agent feedback 1:47 - Capturing sessions with Tapes 3:43 - Labeling expert pushback in Paper Console 5:59 - Turning traces into a review dataset 6:28 - The Twin Review prototype 8:06 - Useful labels and recurring failure patterns 9:32 - Start with skills and evaluations 10:48 - Model adaptation experiments and closing advice
Not rated yetThe Agents Were Fine. Everything Around Them Broke. - Joey Wang, Harvey
· 12 min
After 10,000 agent-written pull requests at Harvey, the agents weren't the problem. Everything around them was. Joey Wang tells three stories from Spectre, Harvey's internal cloud agent platform: letting engineers pick their own agent while one platform owns what happens after, making every action callable so agents can drive other agents, and what broke once they did. Expect an alert storm that spun up hundreds of tasks, a full cluster, a merge queue that couldn't keep up, and the fixes they built in response. Joey Wang, Harvey: https://x.com/ZongZiWang | https://www.harvey.ai
Not rated yet60% of Our Data PRs Now Merge With No Human Review - Chetas Joshi, Robinhood
· 28 min
Coding agents made data pipeline PRs cheap, and human review couldn't keep up. Chetas Joshi shows the hybrid reviewer agent Robinhood built for its data platform: deterministic checks run first, the model judges what scripts can't, and anything risky still goes to a person. Today about 60% of those PRs merge with no human review, saving roughly 132 engineering hours a month. He covers where they draw the line between agent and human, how they evaluate the reviewer itself, and what they got wrong. Chetas Joshi, Robinhood: https://robinhood.com
Not rated yetAn Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
· 38 min
At LinkedIn, AI agents pull together profiles, preferences, career goals and activity into an evolving picture of each member's career. Saloni Gandhi and Ziwei Guo show the closed loop that keeps that agent improving: it evaluates its own outputs, finds failure patterns, reflects on root causes, and proposes targeted fixes, using GEPA to refine its instructions. They also share where they're taking it next, beyond prompts to improving the whole agent harness. Saloni Gandhi, LinkedIn: https://x.com/sal_gandhi | https://www.linkedin.com Ziwei Guo, LinkedIn: https://www.linkedin.com
Not rated yetStop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo
· 1 hr 50 min
Fast answers are easy to get. Answers you can trust with your money are harder. Francesco Galletta and Felippe Felisola Caso build a financial-analyst agent harness live, using one real problem: paying for a wedding. They show the five design principles they use for AI you can trust: shared domain, consistency, evidence, accountability and verification. They also explain why calculations belong in code instead of in the model's head. Note: the speakers work at Monzo but are not presenting on Monzo's behalf. 🎟️ Join us at the next one: AIE CODE Summit → https://aie.codes Speakers Francesco Galletta — LinkedIn: https://www.linkedin.com/in/francescogalletta/ Felippe Felisola Caso — LinkedIn: https://www.linkedin.com/in/felippecaso/ Monzo — https://monzo.com Timestamps 0:00 Intro: Francesco & Felippe 2:54 Using AI for impact, not token-maxing 4:22 Speed isn't the bottleneck. Trust is. 7:25 Components vs. design principles 8:45 The 5 principles of trust 12:13 The use case: budgeting a wedding 17:05 How the principles build on each other 20:40 Principle 1: Shared domain 25:59 Live demo: planning with the agent 30:42 From plan to decision tree 35:48 Why long linear chats fall apart 36:48 Principle 2: Consistency (math belongs in code) 55:50 Principle 3: Evidence (show the receipts) 1:00:17 The "master file" horror story 1:06:42 Principle 4: Accountability 1:20:39 Principle 5: Verification 1:24:16 Golden examples & outside reference points 1:34:05 Q&A: How many harnesses should you build? 1:38:11 Don't over-sharpen the axe 1:44:00 Q&A: Skills vs. a dedicated harness 1:48:31 Takeaway: design for trust on purpose
Not rated yetYour Self-Hosted LLM Needs to Say No - Harshul Jain & Parv Khatri, Audible & SAP
· 52 min
Coding agents fan out, retry and burst, and if they hit an LLM you host yourself, your inference layer will happily accept every request and bury the ones you needed first. Harshul Jain and Parv Khatri make the case for admission control: deciding what to refuse before it reaches the GPU. Drawing on a review of 55 systems papers and eight real production incidents, they share five questions to ask any admission policy before you deploy it, and a checklist you can use right away. Harshul Jain, Audible: https://x.com/hj1393 | https://www.audible.com Parv Khatri, SAP: https://www.sap.com
Not rated yetHow One Founder Manages 20 AI Coding Agents at Once - Ryan Carson, Untangle
· 15 min
Ryan Carson, founder and CEO of Untangle and a four-time founder, shows how his team runs software development with agents in the cloud. One "engineering manager" agent, which he mostly talks to by voice, manages up to 20 builder agents through a ticket board, with an automation that checks every builder every 15 minutes, capped review loops, QA agents and a merge queue manager. The human step he won't skip is watching the video of each feature before it ships. He also explains how Untangle now hires: candidates record themselves shipping a real feature with an agent, an agent ranks the videos, and the top five do a paid trial. Chapters: 0:00 Intro 0:12 What Untangle does 1:02 Managing agents like managing people 1:55 The org chart in the age of agents 2:40 Talking to your manager agent 3:29 One manager, 20 builders 5:05 Review agents and QA 6:24 Watch the video before you ship 8:22 Stop coding locally 8:43 Running it across a team 9:46 The merge queue manager 11:07 How hiring changes 12:24 Ranking candidates with an agent 13:24 A paid trial project 14:50 Wrap-up Ryan Carson https://untangle.us Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetAI Engineer NYC 2026 Closing Remarks - Ksenia Se, Turing Post & swyx, AI Engineer
· 3 min
Ksenia Se of Turing Post and swyx close out three days of AI Engineer NYC 2026, from building agents to changing how financial institutions work. They thank the speakers, sponsors and crew, and share what's next: AIE CODE, AI Engineer Europe in London in February, and the World's Fair in San Francisco next summer. Ksenia Se https://www.turingpost.com swyx Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetYour KV Cache Can Spot Attacks on Your AI Agents - Ari Azarafrooz, Intrinsec AI
· 7 min
Ari Azarafrooz, founder of Intrinsec AI, shows how the KV cache your inference engine already writes can double as a security signal for AI agents. Instead of only guarding inputs and outputs with judge models or a classifier for every failure, he learns a baseline from normal traffic and flags KV pages that look unusual, catching problems like a hidden spoken instruction in a voicemail where there is no text to filter. He covers where this plugs into stacks like vLLM, SGLang, LMCache, llm-d and NVIDIA Dynamo, and the two hard parts in production: models that keep changing and KV data too big to move in real time. Ari Azarafrooz https://intrinsec.ai Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetWill AI Agents Pretend to Be Human to Get Past a Check? - Maty Bohacek, Stanford
· 17 min
Maty Bohacek of Stanford looks at a security check we still rely on: getting on a phone or video call to prove you're human before opening an account, approving a wire or recovering access. With AI avatars that can now join Zoom calls, he tested what agents do when a human check is the only thing between them and their goal. When a voice or video cloning tool was available, many agents used it to impersonate a person, even when told not to, and safety alignment only cut the rate from 65% to 44%. He closes with what needs to happen next across alignment, control, and deepfake detection and provenance. Chapters: 0:00 Intro 0:58 Proving you're human on a call 3:19 When AI controls fail 6:00 Human verification as free AI control 6:41 AI avatars that join video calls 7:51 Experiment: agents blocked by a human check 12:05 How often agents impersonate people 13:47 Can people still tell? 15:18 Where we go from here 17:30 Keep verification safe Maty Bohacek Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetThe One Thing Every AI Agent Framework Agrees On - Mislav Ivanda, Daytona
· 15 min
Mislav Ivanda of Daytona has integrated sandboxes into most of the major agent frameworks, and noticed that frameworks that disagree on almost everything all ended up in the same place. He compares three ways to run agent-generated code: allow lists and approval prompts, embedded runtimes, and a separate sandboxed computer. He explains why only the last one keeps a compromised step from compromising your app, and shows frameworks like CrewAI, AG2 and the OpenAI Agents SDK moving to sandboxes. His advice: models, prompts and even frameworks are easy to swap, so choose where your agent's code runs before you choose a framework. Chapters: 0:00 Intro 0:52 Guess the framework 2:20 Frameworks that agree on nothing 3:42 Where should agent code run? 4:12 Option 1: allow lists and approvals 5:23 Option 2: embedded runtimes 6:46 Option 3: a separate computer 7:54 Frameworks are moving to sandboxes 9:19 How sandbox use has changed 10:33 Which framework should I pick? 12:46 The one choice you can't easily swap 14:09 Frameworks are opinions, boundaries are commitments Mislav Ivanda https://www.daytona.io Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetWhy No One Agrees on What Their AI Tokens Cost - Pratyusha Duvvuri, Cisco
· 18 min
Pratyusha Duvvuri, who led agent observability at Galileo before it became part of Cisco, asked about 100 companies what "tokenomics" means and got 100 different answers. She shares her definition, the observability and interpretability of token usage and cost, and why engineers, engineering managers and finance teams each need something different from it. Then she demos the tokenomics dashboard her team took from first customer call to GA in 10 weeks: spend by person and team, adoption trends, forecasting and a finance view. She also covers the hard parts, like custom provider pricing and bills that change after the fact. Chapters: 0:00 Intro 0:12 What does tokenomics mean to you? 1:49 Three trends in AI spend 4:42 Three kinds of users 6:46 Why provider dashboards aren't enough 8:34 A definition of tokenomics 9:18 What you need to get started 10:51 Demo: the tokenomics dashboard 14:35 The finance view 15:48 What's coming next 16:44 Challenges building it Pratyusha Duvvuri Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetTraining Your Own Agent Models With RL Is Now Practical - Jason Mancuso, Modal
· 20 min
Jason Mancuso of Modal argues that treating models as swappable backends is a losing strategy, and that teams with agents in production already have what they need for reinforcement learning: a harness, evals and real traces. He explains why RL generalizes better than fine-tuning on recorded traces, then walks through a stable recipe for asynchronous RL built on GRPO with masked importance sampling, score centering and top-p sampling. He shows how sparse weight updates let Modal spread rollouts across clouds, GPU types and quantization levels while training stays stable and rollouts get 15% cheaper. Chapters: 0:00 Intro 0:20 Models aren't static backends 1:09 Why teams fine-tune 3:35 Why RL instead of SFT 4:46 You already have what you need 6:51 How reinforcement learning works 7:41 Async RL and staleness 9:38 A stable RL recipe 12:38 Why RL on Modal is different 14:55 Sparse weight updates 16:21 Mixing GPU types for rollouts 17:56 15% cheaper rollouts 19:16 Own your stack Jason Mancuso https://modal.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetRight Answer, Wrong Reasons: AI Forecasts Costco - Hassan El Mghari & Omar El Mghari, Together AI
· 15 min
Hassan El Mghari of Together AI and his father Omar, a former CFO and head of internal audit, test whether an open model can forecast a company's profits. They give GPT-OSS-120B only Costco's public filings from before its 2024 membership fee increase, have it write the financial model as structured formulas, and run 16 cheap attempts in parallel. The forecast lands within 4.4% of Costco's actual results, until an audit of the assumptions shows two errors that happened to cancel out. Their takeaway: when you build agents for specialized work, put an expert in the loop and turn their checks into code. Chapters: 0:00 Intro 0:30 Can an AI forecast Costco's profits? 0:49 Meet the father-son team 2:21 Spoiler: 4.4% off 3:03 Why membership fees matter to Costco 4:12 Using models that didn't already know the answer 5:03 Turning 214 filings into 3,000 facts 5:50 Have models write formulas, not do the math 7:18 Checks and 16 parallel runs 8:41 The forecast 9:53 An auditor looks at the assumptions 11:39 Two errors that canceled out 13:29 Three lines of defense 14:48 Put an expert in the loop Hassan El Mghari Omar El Mghari https://www.together.ai Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetHow to Build a Software Factory That Improves Itself - Zach Lloyd, Warp
· 22 min
Zach Lloyd, founder and CEO of Warp and former engineering lead on Google Docs, explains why companies are moving from engineers prompting agents on their laptops to sandboxed cloud software factories. He makes the case for treating your factory as infrastructure that's open, works with your existing tools and is defined in code. He demos Warp Factories: visibility into every agent run, LLM-judged scores, a self-improvement loop that proposes fixes to the factory itself, and benchmarks built from your own past tasks to compare cost and quality across models. Chapters: 0:00 Intro 1:12 From interactive prompting to automated development 2:21 Why local agents hit limits 3:08 What a cloud software factory is 4:03 Treat your factory as infrastructure 5:15 Own your data and compute 6:01 Measure on your own data 6:55 Product engineers and factory engineers 7:51 Demo: Warp Factories 10:33 Defining the factory in code 12:01 Scoring agent runs 13:21 The self-improvement loop 14:42 Benchmarks on your own tasks 17:15 The factory stack 17:57 Q&A Zach Lloyd https://www.warp.dev Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetStop Doing Agent Archaeology on Your Team's Code - Richard Feldman, Delta
· 19 min
Richard Feldman of Delta explains why handing off agent-driven work to a teammate loses so much: you get the branch and maybe a pasted transcript, but not the reasoning, the dead ends or the decisions made along the way. He demos Delta, a tool where the agent conversation, code changes, reviews and teammates all live in one shared thread. It shows @-mentioning a coworker into a thread, line-by-line code review that goes straight to the agent, live terminal sync, and instant local checkouts of a teammate's work. He finishes with audience questions on scale, on-prem deployment, CI and catching up on long threads. Chapters: 0:00 Intro 0:25 The Monday morning handoff 2:31 How teams lose context with agents 3:26 Agent archaeology 4:40 Humans and agents collaborating together 7:21 Demo: Delta 8:28 Bringing a coworker into the thread 9:49 Code review inside the thread 11:05 Live terminal sync and shared checkouts 13:00 Wrapping up 14:32 Q&A Richard Feldman https://delta.dev Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetTeach Your AI Agent to Build Products Everyone Can Use - Jemima Abu, CAIS
· 14 min
Jemima Abu, a product engineer and certified accessibility professional, explains why AI agents copy the accessibility gaps in the code they learned from, and why 96% of top websites still have accessibility issues. In an informal experiment, giving her agent an accessibility skill made the same build faster and cheaper. She covers what inaccessible code costs you in users, money and legal risk, then shares a four-step approach: define requirements, audit your codebase, use a design system, and make your agents follow the rules every time. She also shares a lightweight accessibility skill you can start with. Chapters: 0:00 Intro 0:12 Unlock your phone with your eyes closed 1:34 What accessibility means 2:04 Why 96% of websites fall short 2:44 An experiment with an accessibility skill 4:11 Your AI agent is a toddler 5:12 What inaccessible code costs you 9:46 Getting agents to write accessible code 12:03 A lightweight accessibility skill 13:27 Q&A Jemima Abu Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetCan AI Agents Beat Wall Street at Predicting Earnings? - Ashwin Paranjape, Samaya AI
· 15 min
Ashwin Paranjape, founding AI lead at Samaya AI, explains why stock prediction is too noisy to learn from, and why earnings prediction makes a better test for financial agents: it has hard numbers, an expert baseline and an explanation for every surprise. He walks through a point-in-time environment that blocks any look at future data, then results across more than 450 earnings reports where the latest frontier models start to beat analyst consensus. He shows where agents win and miss, and why live data matters more than anything else. Chapters: 0:00 Intro 0:39 About Ashwin 1:52 Finance's big unsolved problem 2:19 Why not predict stocks? 3:18 Earnings prediction as a task 4:32 A point-in-time environment 6:31 Results: AI vs. consensus 9:30 Where agents win 10:35 Where agents miss 11:46 Why live data matters most 13:09 Expert guidance and hiding consensus 14:11 What's next Ashwin Paranjape https://samaya.ai Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yetWhat Your Coding Agent Can't See in the Code - Tom Moor, Linear
· 21 min
Tom Moor of Linear argues that writing code is no longer the bottleneck. The harder problem is keeping a product coherent as agents ship more and more of it. He walks through how Linear carries context from intake to merge: a triage agent that enriches each issue with customer and incident context, a review agent that checks changes against past decisions, reverts and regressions, and a daily automation that turns review feedback into lessons in the codebase. He shares real examples where that history caught problems a standard code reviewer would have missed. Chapters: 0:00 Intro 0:37 What Linear is 1:48 How fast software engineering changed 5:16 Context moves to the outer loop 6:04 Code isn't the bottleneck anymore 6:57 What agents miss in the code 8:57 Automating intake with a triage agent 12:16 Building with verification 13:40 Context-aware code review 15:53 Review examples from Linear 18:16 Turning feedback into lessons 20:04 What carrying context buys you Tom Moor https://linear.app Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!
Not rated yet