Summaries of every talk on the AI Engineer YouTube channel.https://aietalks.comen-gbhttps://aietalks.com/talks/agent-spending-without-controlshttps://aietalks.com/talks/agent-spending-without-controlsRodrigo Coelho and Pranav Maheshwari argue that agentic commerce needs two pieces of infrastructure: a way for agents to pay for useful tools and controls that prevent unsafe transactions. Coelho gives the history, including Edge & Node's query micropayment work for The Graph in 2021 and its later collaboration on x402. Maheshwari then demonstrates the difference between an agent with...Tue, 01 Sep 2026 00:00:00 GMTenterpriseguardrailsmcptool-usehttps://aietalks.com/talks/beyond-the-lethal-trifecta-agentic-commerce-on-the-open-internethttps://aietalks.com/talks/beyond-the-lethal-trifecta-agentic-commerce-on-the-open-internetDavid Levine traces his idea of agentic commerce back to LambdaMOO, a text-based virtual world he joined in 1993. He says its sense of reality came from governance, technology, economics, and culture working together. Platforms later broke apart that kind of community by extracting value through advertising and engagement algorithms. Levine argues that agents face a similar problem today. Private...Tue, 01 Sep 2026 00:00:00 GMTagentssecurityhttps://aietalks.com/talks/multimodal-collaborative-agents-for-next-gen-commercehttps://aietalks.com/talks/multimodal-collaborative-agents-for-next-gen-commerceNidhi Kaushik Vyas presents a framework for commerce agents that work with incomplete, subjective intent. She argues that users often arrive with a vibe rather than a well-formed query, so the agent must build a working state from conversation history, personal context, hard constraints, and reference images. It should attach confidence to inferred preferences and identify information that needs real-time...Tue, 01 Sep 2026 00:00:00 GMTagentsevalsmultimodalplanninghttps://aietalks.com/talks/teaching-agents-to-payhttps://aietalks.com/talks/teaching-agents-to-payAnna Spysz builds a shopping agent to find headphones for recording, mixing, and mastering music. The example introduces the infrastructure needed for an agent to transact, including the Universal Commerce Protocol, merchant capability manifests, structured catalogs, policy data, and logs that record why products were recommended. A local record shop becomes visible to the agent after publishing this information in...Tue, 01 Sep 2026 00:00:00 GMTagentsguardrailsprivacystructured-outputshttps://aietalks.com/talks/the-end-of-the-static-screen-architecting-intent-driven-uxhttps://aietalks.com/talks/the-end-of-the-static-screen-architecting-intent-driven-uxGus Iwanaga argues that software still makes people adapt to its fixed interfaces, even though AI can generate more personalized experiences. He demonstrates why unconstrained generative UX is difficult to ship: the same sales-report request produced different KPI cards, charts, text, and date ranges on four attempts. His team chose a middle path between fixed components and fully generated HTML....Tue, 01 Sep 2026 00:00:00 GMTagent-skillsagentsproduct-strategystructured-outputstool-usehttps://aietalks.com/talks/when-ai-agents-pay-and-sellers-monetize-building-x402-apps-on-awshttps://aietalks.com/talks/when-ai-agents-pay-and-sellers-monetize-building-x402-apps-on-awsAnil Nadiminti argues that web payment systems built around human subscriptions and card transactions do not fit autonomous agents. Agents may need to pay a fraction of a cent for one API call, while card rails impose a 25-cent minimum plus a percentage fee. When agents meet paywalls, humans must intervene, which breaks the autonomous workflow. Sellers face a choice...Tue, 01 Sep 2026 00:00:00 GMTagentsenterprisesecuritytool-usehttps://aietalks.com/talks/why-your-ai-agent-needs-a-wallet-usdc-and-nanopaymentshttps://aietalks.com/talks/why-your-ai-agent-needs-a-wallet-usdc-and-nanopaymentsHarshal Bhangale argues that payment is a practical limit on AI agents. Agents can research across many sources, but they stop at paywalls because existing payment flows assume a human can sign up, enter card details, or manage API keys. Cards are poorly suited to frequent one-cent payments because a fee near 3% is too large. In his demo, two...Tue, 01 Sep 2026 00:00:00 GMTagentsguardrailshttps://aietalks.com/talks/x402-isnt-good-yethttps://aietalks.com/talks/x402-isnt-good-yetJan Curn presents x402 as an exciting protocol that still has serious implementation problems. Apify operates a marketplace with about 45,000 tools, so it needs payments that work for agents running anything from simple API calls to jobs lasting hours. The main problem is the gap between signature verification and blockchain settlement. A client can create many signatures against the...Tue, 01 Sep 2026 00:00:00 GMTenterprisetool-usehttps://aietalks.com/talks/your-agent-just-authorized-whathttps://aietalks.com/talks/your-agent-just-authorized-whatJay Mok and Ben Coumes present a framework for deciding how agents should be authorized. Every system needs to answer three questions: did a human authorize the action, is it allowed now within the stated scope, and can the authorization be proven later? Their answer depends on the stakes and on whether the parties know each other. A coding agent...Tue, 01 Sep 2026 00:00:00 GMTagentshuman-in-the-loopprivacysecurityhttps://aietalks.com/talks/sota-generative-media-panelhttps://aietalks.com/talks/sota-generative-media-panelThe panel covers Google DeepMind's recent generative media work, including Nano Banana 2 Lite and the Gemini Omni Flash APIs. Nicole Brichtova describes image-to-video workflows, editing, translation, marketing, and education use cases. The discussion then turns to the architecture of multimodal systems. Shane Gu argues that language is a useful backbone for reasoning and generalization, while Dumitru Erhan says video...Sun, 30 Aug 2026 00:00:00 GMTevalsmultimodalsynthetic-datavideohttps://aietalks.com/talks/agentic-sites-building-hyper-personalized-websiteshttps://aietalks.com/talks/agentic-sites-building-hyper-personalized-websitesCarlos Sanchez presents Adobe's approach to agentic sites, where a website adapts itself to a visitor's intent in real time. The system records browsing signals, groups visitors into intent or persona types, and uses those signals to change selected blocks such as the hero, product list, navigation, blog feed, and calls to action. The whole site is used as a...Sat, 29 Aug 2026 00:00:00 GMTlatencyproduct-strategyraghttps://aietalks.com/talks/agents-are-where-microservices-were-in-2015https://aietalks.com/talks/agents-are-where-microservices-were-in-2015Roberto Milev and Uday Kanagala compare the current agent wave with the microservices wave. Their advice is to learn the basic unit before adding orchestration. At Navan, that means running a single master agent that progressively loads reusable skills. A skill contains domain instructions, setup, and tool execution, so it can be tested and reused independently. The talk also covers...Sat, 29 Aug 2026 00:00:00 GMTagentsevalsguardrailsobservabilityhttps://aietalks.com/talks/ai-agents-are-just-distributed-systems-nowhttps://aietalks.com/talks/ai-agents-are-just-distributed-systems-nowSalman Munaf argues that tool-using agents should be designed as distributed systems. A text-only model could produce a bad answer, but an agent can call an API, update a database, send an email, or issue a refund. That creates network failures, stale state, duplicate requests, partial completion, retry storms, and unsafe permissions. Munaf describes the agent as a probabilistic coordinator...Sat, 29 Aug 2026 00:00:00 GMTguardrailsobservabilityreliabilitytool-usehttps://aietalks.com/talks/from-tokenmaxxing-to-trusted-throughputhttps://aietalks.com/talks/from-tokenmaxxing-to-trusted-throughputMingsheng Hong argues that engineering teams should stop treating token usage as a target. A dashboard can reveal adoption gaps, sudden usage bursts, or unusual differences between teams, but high usage is not proof of value. Ironclad first measured lines of code, then open pull requests, then merged pull requests. Its current measure weights merged pull requests by an AI-generated...Sat, 29 Aug 2026 00:00:00 GMTcostdeveloper-experienceobservabilityreliabilityhttps://aietalks.com/talks/tell-the-robot-what-you-wanthttps://aietalks.com/talks/tell-the-robot-what-you-wantSandhya Subramani demonstrates Scout, a four-legged robot with a Raspberry Pi, a SIM card, and a 4G connection. Scout already has preset movement policies, but an agent layer lets it interpret natural-language requests and choose which policy to call. Subramani uses AWS's open-source Strands Agents framework to connect the robot as a hardware tool in about five lines of code....Sat, 29 Aug 2026 00:00:00 GMTagentscomputer-useedgetool-usehttps://aietalks.com/talks/the-half-life-of-agent-infrastructurehttps://aietalks.com/talks/the-half-life-of-agent-infrastructureBen Kus argues that infrastructure advice built around long-lived systems does not fit AI agents. Databases, identity controls, storage, and engineering practices can remain useful for years, while models, agent patterns, harnesses, and retrieval methods can change within months. He illustrates this through a series of shifts, from custom and fine-tuned models to frontier models, open-weight models, adaptive model selection,...Sat, 29 Aug 2026 00:00:00 GMTagentsengineering-cultureevalsraghttps://aietalks.com/talks/the-signal-layer-what-to-build-when-anything-can-be-builthttps://aietalks.com/talks/the-signal-layer-what-to-build-when-anything-can-be-builtLena Hall argues that AI has made average software and content cheap to produce, because competitors can point similarly capable systems at the same questions and receive similar answers. The advantage therefore moves to deciding what to build and how to preserve the reason it should be chosen. She separates this into knowing the signal, then shipping it without distortion....Sat, 29 Aug 2026 00:00:00 GMTproduct-strategyhttps://aietalks.com/talks/tribal-dungeons-of-global-shipping-ai-agents-at-global-scalehttps://aietalks.com/talks/tribal-dungeons-of-global-shipping-ai-agents-at-global-scaleDmitry Buykin describes agent work in global shipping, where a shipment is really an orchestration of many parallel state machines. The easy cases are already automated, while exceptions remain expensive because they cross incomplete legacy systems and vary by country. Maersk's existing procedures were screenshots that showed people what to click. Buykin's team translated them into executable procedures with preconditions,...Sat, 29 Aug 2026 00:00:00 GMTagentsguardrailsobservabilityworkflowshttps://aietalks.com/talks/which-ai-startups-actually-land-enterprise-contractshttps://aietalks.com/talks/which-ai-startups-actually-land-enterprise-contractsBrian Lewis explains enterprise AI buying from the perspective of a buyer at Millennium, while speaking personally rather than for his employer. For one internal pain point, he may review 10 to 15 startups, book two or three demos, run zero or one pilot, and sign about one contract for every four pilots. The talk examines where the other opportunities...Sat, 29 Aug 2026 00:00:00 GMTenterpriseproduct-strategysecurityteam-adoptionhttps://aietalks.com/talks/ai-evals-for-cross-functional-teamshttps://aietalks.com/talks/ai-evals-for-cross-functional-teamsDoorDash's GenAI platform team began with the assumption that evals were an engineering concern. Different product groups needed different forms of evaluation, including session-level judgments, scaled human review, and trajectory-based evaluation. The team therefore moved from UI-first to API-first, then toward workflow-first tools that let strategy and operations staff run work themselves. The talk describes an eval process shared across...Fri, 28 Aug 2026 00:00:00 GMTevalshuman-in-the-loopobservabilityworkflowshttps://aietalks.com/talks/ai-native-organisations-run-on-skills-how-to-structure-and-scale-themhttps://aietalks.com/talks/ai-native-organisations-run-on-skills-how-to-structure-and-scale-themImad Touil argues that skills are where an AI-native organization's practical know-how becomes executable. He places them inside a larger agentic stack with workflows, hooks, MCP servers, subagents, and context management, then explains why those other components do not capture enough reusable structure. Skills can define specialized tasks, compose across workflows and harnesses, reduce context usage through progressive disclosure, and...Fri, 28 Aug 2026 00:00:00 GMTagent-skillsagentssecurityteam-adoptionworkflowshttps://aietalks.com/talks/building-the-engine-while-flying-the-plane-launching-the-figma-mcp-serverhttps://aietalks.com/talks/building-the-engine-while-flying-the-plane-launching-the-figma-mcp-serverJesse Lumarie describes how Figma built and shipped its first MCP server in about three months while the protocol and client ecosystem were changing. He began the work as a one-day-a-week project after seeing an internal demo and wanting non-designers to use Figma through AI tools. The team had to handle a deprecated transport, uneven client support, and incomplete implementations...Fri, 28 Aug 2026 00:00:00 GMTcode-generationenterpriseevalsmcphttps://aietalks.com/talks/building-ureview-ubers-multi-agent-code-review-enginehttps://aietalks.com/talks/building-ureview-ubers-multi-agent-code-review-engineWill Bond and Ameya Ketkar explain why Uber built uReview instead of buying an automated code review product. Uber needed support for Phabricator, the same review rules in human and agent workflows, team-owned customizations, and routing based on change risk and complexity. The system accepts reviews from GitHub, Phabricator, and the agent loop, then rates, filters, categorizes, and deduplicates comments....Fri, 28 Aug 2026 00:00:00 GMTagentscoding-agentsevalsmulti-agentobservabilityhttps://aietalks.com/talks/from-ai-assisted-to-ai-native-building-a-frontier-development-teamhttps://aietalks.com/talks/from-ai-assisted-to-ai-native-building-a-frontier-development-teamClare Liguori describes a shift from chat-based coding assistance to what Amazon calls frontier development. Frontier developers write only 1 to 2 percent of their code themselves, run agents for hours, and keep several agents working in parallel. In Amazon's pilot with 50 ordinary teams using existing codebases, half saw less than a 3x increase in deployment velocity, while the...Fri, 28 Aug 2026 00:00:00 GMTagent-skillsagentscoding-agentscontext-engineeringtestinghttps://aietalks.com/talks/how-do-you-diffuse-ai-into-the-real-worldhttps://aietalks.com/talks/how-do-you-diffuse-ai-into-the-real-worldVarun Shenoy argues that the hard part of AI is no longer demonstrating what models can do. It is getting them to complete economically useful work inside real companies. He compares the current moment with electricity, which needed new factory equipment and retraining before it changed production. Long Lake approaches this problem by owning the services businesses where it deploys...Fri, 28 Aug 2026 00:00:00 GMTagentsenterpriseevalsteam-adoptionhttps://aietalks.com/talks/how-to-avoid-disaster-when-vibe-coding-a-billing-enginehttps://aietalks.com/talks/how-to-avoid-disaster-when-vibe-coding-a-billing-engineAndrew Garvin demonstrates a Stripe Projects workflow that uses a coding agent to create a Metronome billing sandbox from a natural-language request. His example copies Lovable's prepaid credit model, including auto recharge, separately scoped credit pools, usage, and a draft invoice. The demonstration depends on portable skills files that explain Metronome's API and verbose errors that let the agent recover...Fri, 28 Aug 2026 00:00:00 GMTagent-skillsagentscoding-agentshuman-in-the-looptestinghttps://aietalks.com/talks/how-to-get-your-org-to-adopt-coding-agents-without-shipping-garbagehttps://aietalks.com/talks/how-to-get-your-org-to-adopt-coding-agents-without-shipping-garbageEyal Blum describes AI adoption at Figma as a three-act process. Engineers first find small tasks where agents work well, then lose trust when they apply the same approach to larger problems, and finally learn to use better context, prompts, and guardrails. Adoption is uneven, so teams with different levels of confidence still have to work together. Blum says the...Fri, 28 Aug 2026 00:00:00 GMTagentscoding-agentsguardrailsteam-adoptiontestinghttps://aietalks.com/talks/productionizing-llm-gateways-architecture-tradeoffs-and-hard-lessonshttps://aietalks.com/talks/productionizing-llm-gateways-architecture-tradeoffs-and-hard-lessonsKanish Manuja explains the engineering choices behind production LLM gateways. A gateway sits between applications and model providers, handling routing, authentication, fallbacks, rate limits, and governance. During failures, teams cannot maximize availability, latency, guardrails, and cost at the same time. Manuja recommends per-request fallback instead of blindly retrying a slow and expensive provider, with careful capacity planning for the fallback...Fri, 28 Aug 2026 00:00:00 GMTcostguardrailslatencyreliabilityhttps://aietalks.com/talks/your-code-has-bugs-lean4-has-proofs-formal-verification-for-engineershttps://aietalks.com/talks/your-code-has-bugs-lean4-has-proofs-formal-verification-for-engineersVarun Pant argues that coding agents have made ordinary software checks insufficient. Model-based grading is probabilistic, tests cover selected inputs, and human review cannot keep pace with agent-generated pull requests. Formal verification starts with a specification of what correct means, then proves that the implementation satisfies it for every input. Pant assigns humans responsibility for validating the specification and machines...Fri, 28 Aug 2026 00:00:00 GMTagentscoding-agentstestinghttps://aietalks.com/talks/can-llms-write-fast-multi-gpu-kernelshttps://aietalks.com/talks/can-llms-write-fast-multi-gpu-kernelsSimran Arora explains why multi-GPU communication has become a larger performance problem as GPU compute has advanced faster than the links between GPUs. She introduces the hierarchy from PCIe to NVLink and NVSwitch, then describes the choices kernel authors must make around copy engines, tensor memory acceleration, register-level transfers, and the placement of communication relative to computation. Together AI captured...Thu, 27 Aug 2026 00:00:00 GMTbenchmarkscode-generationgpushttps://aietalks.com/talks/how-anthropic-builds-lessons-from-labshttps://aietalks.com/talks/how-anthropic-builds-lessons-from-labsMike Krieger describes how his work changed after moving from Anthropic's chief product officer role into an individual contributor position at Labs. He now gives Claude a goal and lets it work through the details, rather than specifying every engineering step. He argues that people should ask more of AI systems, especially as they gain access to tools and environments....Thu, 27 Aug 2026 00:00:00 GMTagentscoding-agentsengineering-cultureproduct-strategyreliabilityhttps://aietalks.com/talks/how-to-generate-mergeable-code-with-a-context-enginehttps://aietalks.com/talks/how-to-generate-mergeable-code-with-a-context-enginePeter Werry argues that agents need more than access to a codebase or a large context window. Before agents, engineers carried organizational context by reading code, searching discussions, and remembering decisions from incidents and past work. Agents reset this knowledge for every task, so they often find one plausible answer and stop searching. Werry calls this failure "satisfaction of search,"...Thu, 27 Aug 2026 00:00:00 GMTagentscode-generationcoding-agentscontext-engineeringplanninghttps://aietalks.com/talks/kv-cache-aware-routing-and-p-d-disaggregation-on-kuberneteshttps://aietalks.com/talks/kv-cache-aware-routing-and-p-d-disaggregation-on-kubernetesYuchen Fama and Ashish Kamra describe two ways to run inference for agentic workloads on Kubernetes. The first is KV cache-aware routing. Agent sessions can run for thousands of turns, reuse system prompts, and produce cache hit rates above 90 percent. Since cached and uncached tokens can have a 10x cost difference, llm-d routes requests toward pods with the right...Thu, 27 Aug 2026 00:00:00 GMTcachingdeploymentinferencelatencyhttps://aietalks.com/talks/the-agentic-commerce-stackhttps://aietalks.com/talks/the-agentic-commerce-stackAhnaf Prio describes what Best Buy has learned while working on agentic commerce. Shopping already makes up about 45% of sessions on major AI assistants, but the first browser-based agents struggled with screenshots, DOM changes, slow form filling, and merchant fraud checks. The emerging approach uses merchant APIs and product feeds. MCP gives an agent access to tools, A2A lets...Thu, 27 Aug 2026 00:00:00 GMTagentsevalsmcpmulti-agenthttps://aietalks.com/talks/ai-in-gtm-at-notionhttps://aietalks.com/talks/ai-in-gtm-at-notionFlora Liu describes Notion's effort to unify self-serve growth and sales-assisted GTM into one decisioning system. Customer data previously lived across Salesforce, Gong, Outreach, ZoomInfo, Snowflake, and years of notes and meeting documents. The team found that stale data, bad account mappings, and unreadable notes could make automation act on the wrong facts. Their architecture has four layers: a trusted...Wed, 26 Aug 2026 00:00:00 GMTagentsdata-pipelineshuman-in-the-loopworkflowshttps://aietalks.com/talks/building-gtm-ai-agents-lessons-from-deploying-to-6-000-usershttps://aietalks.com/talks/building-gtm-ai-agents-lessons-from-deploying-to-6-000-usersSait Izmit describes Snowflake's internal go-to-market assistant, which launched in September and now serves about 6,000 users. His approach starts with quality. Before connecting more data, his team wrote 150 questions from the sales process and used them to expose gaps. The first test reached 50 percent accuracy, leading the team to focus on answering fewer questions well. The assistant...Wed, 26 Aug 2026 00:00:00 GMTagentsevalsteam-adoptionworkflowshttps://aietalks.com/talks/gtm-engineering-the-technical-bitshttps://aietalks.com/talks/gtm-engineering-the-technical-bitsEverett Berry describes GTM engineering as the technical work needed to let go-to-market teams ship data, automations, and campaigns at an engineering-like pace. He focuses on four problems. Data must track companies that are constantly changing, while combining incomplete information from many vendors and keeping update costs under control. Orchestration must connect ten to thirty tools with different data volumes,...Wed, 26 Aug 2026 00:00:00 GMTagentsdata-pipelineshuman-in-the-loopworkflowshttps://aietalks.com/talks/how-ai-agents-let-gtm-teams-scalehttps://aietalks.com/talks/how-ai-agents-let-gtm-teams-scaleJustin Joyce argues that traditional go-to-market operations break down as teams grow. Operations staff rebuild analysis in spreadsheets or maintain dashboards that cannot answer every request. Salespeople also lose time switching between prospects, customers, and adoption work while trying to gather the right context. Joyce describes three responses at Cloudflare: scale analysis with skill files that combine business knowledge and...Wed, 26 Aug 2026 00:00:00 GMTagentsenterprisemulti-agentworkflowshttps://aietalks.com/talks/how-we-got-llms-to-recommend-our-open-source-libraryhttps://aietalks.com/talks/how-we-got-llms-to-recommend-our-open-source-libraryChristopher Burns explains how Inth made its open source consent banner library, c15t, easier for language models and coding agents to understand. The work started with good developer experience and expanded into many small documentation changes. Burns describes a framework-neutral pipeline that generates agent-facing files from MDX, while recommending that llms.txt itself be written by hand. He also explains why...Wed, 26 Aug 2026 00:00:00 GMTagentscoding-agentsdeveloper-experiencemcphttps://aietalks.com/talks/knowledge-systems-the-new-gtm-stackhttps://aietalks.com/talks/knowledge-systems-the-new-gtm-stackJeffrey Wang argues that go-to-market is an engineering problem because its work is mostly about collecting, organizing, and acting on data. Exa maintains a live model of its customers, prospects, product usage, and the wider web. Its ICP dashboard classifies nearly every company in its target market and attaches account data such as anticipated spend. Request Lens alerts the team...Wed, 26 Aug 2026 00:00:00 GMTagentsdata-pipelinesenterprisemcphttps://aietalks.com/talks/reverse-engineering-the-ai-buyerhttps://aietalks.com/talks/reverse-engineering-the-ai-buyerAliisa Rosenthal uses OpenAI's early enterprise experience to explain how AI companies should build their go-to-market motion. OpenAI had strong demand after ChatGPT launched, but it spent nine months without enterprise features, then built an expensive, high-end product for large companies. When self-serve arrived four months later, it grew faster and competed with the enterprise business. Rosenthal's advice is to...Wed, 26 Aug 2026 00:00:00 GMTenterpriseproduct-strategyworkflowshttps://aietalks.com/talks/the-building-blocks-of-gtm-orchestrationhttps://aietalks.com/talks/the-building-blocks-of-gtm-orchestrationArman Vaziri describes GTM orchestration as the ability to describe a playbook, experiment, or campaign and distribute its execution across channels. He argues that the hard part is usually after the idea: finding the audience, preparing the material, coordinating teams, and getting work done without waiting months. Ramp built an internal customer data platform that combines CRM, product, enrichment, buying-signal,...Wed, 26 Aug 2026 00:00:00 GMTagent-skillsagentsdata-pipelineshuman-in-the-loopworkflowshttps://aietalks.com/talks/the-death-of-developer-advocateshttps://aietalks.com/talks/the-death-of-developer-advocatesStephanie Jarmak argues that developer relations is changing because developers now work with agents, and agents have become users and recommenders of software tools. An agent reads documentation, calls APIs, encounters errors, and decides what to suggest to a developer. Jarmak describes CodeScaleBench, a benchmark built from hundreds of software lifecycle tasks, which she used to compare agents with and...Wed, 26 Aug 2026 00:00:00 GMTagentsdeveloper-experienceevalsmcphttps://aietalks.com/talks/the-missing-layer-in-agentic-aihttps://aietalks.com/talks/the-missing-layer-in-agentic-aiGiedrius Šteimantas uses a personal shopping agent to explain the infrastructure layer that agentic systems need when they work on the open web. The original system used browser automation for discovery, product verification, and checkout. It sometimes returned CAPTCHAs, spent tokens processing invalid pages, and found products that disappeared at checkout because stock and availability depended on location. Šteimantas applies...Wed, 26 Aug 2026 00:00:00 GMTagentscomputer-usecostmcphttps://aietalks.com/talks/einstein-arena-harnessing-collective-agent-intelligence-for-open-sciencehttps://aietalks.com/talks/einstein-arena-harnessing-collective-agent-intelligence-for-open-scienceJames Zou argues that agent systems should be designed around environments rather than fixed workflows. A workflow tells an agent which steps, tools, and prompts to follow. An environment defines the problems, resources, incentives, guardrails, and feedback available to the agent while leaving its method open. Einstein Arena applies this idea to open scientific problems. Agents enter by solving a...Tue, 25 Aug 2026 00:00:00 GMTagentsbenchmarksmulti-agentopen-modelsreinforcement-learninghttps://aietalks.com/talks/agent-frameworks-considered-harmfulhttps://aietalks.com/talks/agent-frameworks-considered-harmfulRémi Louf describes taking two weeks away from running .txt to build a background-agent system for his own morning routine. The system processes market news, CRM and project updates, voice notes, and other events, then posts a daily brief to Slack. He started with markdown agent definitions and cron jobs, but failures quickly exposed the need for a runtime. The...Sat, 22 Aug 2026 00:00:00 GMTagentsobservabilitystructured-outputsworkflowshttps://aietalks.com/talks/finops-for-ai-agents-who-spent-all-the-tokenshttps://aietalks.com/talks/finops-for-ai-agents-who-spent-all-the-tokensTisha Chawla and Susheem Koul describe Token Ops, a control plane for managing the cost of AI agent runs. They argue that SaaS and cloud systems gained usage caps, autoscaling, and related controls, while agent systems still lack controls where code calls a model. The design attributes model-call cost to runs and user-defined dimensions, records it in a ledger, and...Sat, 22 Aug 2026 00:00:00 GMTagentscostguardrailsobservabilityhttps://aietalks.com/talks/give-the-agent-a-budget-not-a-tokenhttps://aietalks.com/talks/give-the-agent-a-budget-not-a-tokenSachin Malhotra argues that token scopes are too blunt for agents working in production. A token answers whether an operation is allowed, but it does not say how often the agent can perform it, how much damage it can cause, whether the action can be undone, or who will notice. He proposes asymmetric verbs, rate limits that refill, trip wires...Sat, 22 Aug 2026 00:00:00 GMTagentsguardrailsreliabilitysecurityhttps://aietalks.com/talks/preferences-over-benchmarks-model-routinghttps://aietalks.com/talks/preferences-over-benchmarks-model-routingArchana Kamath and Tyler Gillam argue that choosing one model by its position on a public benchmark misses the conditions that determine whether it fits a real request. Classification may work well with a small open model, code completion may need a faster model, and code review or security work may justify a frontier model. Their Inference Light Router lets...Sat, 22 Aug 2026 00:00:00 GMTcostevalsinferencelatencyhttps://aietalks.com/talks/the-agent-behind-the-curtain-building-the-oz-cloud-agent-platformhttps://aietalks.com/talks/the-agent-behind-the-curtain-building-the-oz-cloud-agent-platformSafia Abdalla explains how Warp built a cloud platform for agents around a simple principle: the platform should absorb complexity before it reaches the user. Agents need isolated sandboxes, so Warp supports both managed infrastructure and self-hosted environments. Developers can bring different harnesses while conversation state and artifacts follow a consistent structure. Since engineering work rarely fits one prompt, agents...Sat, 22 Aug 2026 00:00:00 GMTagentsdeploymentdeveloper-experiencemulti-agent