AgenticBit

Netflix has written an article about their new LLM Ranker (GenRec) based upon their own internal Foundation Model. I went through the information in depth because the statements made by people on Social Media are much larger than what the article states. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲 𝗮𝗿𝘁𝗶𝗰𝗹𝗲 𝘁𝗿𝘂𝗹𝘆 𝗮𝗰𝗰𝗼𝗺𝗽𝗹𝗶𝘀𝗵𝗲𝘀: The old ranker was built on thousands of manually created features. GenRec takes a member’s watch history and context and transforms them into natural language, then feeds that text into the model to score the entirety of the catalog using a catalog-aware scoring head. To minimize costs, GenRec serves using only prefill inference and context compaction. 𝗥𝗲𝗽𝗼𝗿𝘁𝗲𝗱 𝗿𝗲𝘀𝘂𝗹𝘁𝘀: GenRec showed statistically significant improvements over a long-standing production ranker in both offline and online metrics. GenRec also used approximately 40 times fewer labeled training samples than the production ranker. Offline MRR was improved by approximately 1.6%. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲 𝗮𝗿𝘁𝗶𝗰𝗹𝗲 𝗱𝗼𝗲𝘀 𝗻𝗼𝘁 𝘀𝘁𝗮𝘁𝗲: The article doesn’t state that Netflix will no longer use the old system. The authors describe GenRec as being the first step towards an LLM-Native Stack. Evidence to date only includes batch compute surfaces, supervised reward weighting, and moderate model size. Real Time Ranking and Reinforcement Learning Alignment are listed as unanswered questions. 𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗥𝗲𝗴𝘂𝗹𝗮𝘁𝗲𝗱 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗲𝘀: The Engineering Shift is from Feature Engineering to Context Engineering. The Work is now about determining what Text goes into the Prompt, How History Gets Compacted, and What Reward Signals Help Shape Post Training. From a Governance Seat, This Changes What You Review. A Feature Store Contains Columns You Can Inspect, Test for Drift, and Trace to a Source. A Verbalized User History is a Paragraph. Reviewers Need to Know What Was Included, What Was Dropped During Compaction, and Whether Sensitive Attributes End Up in the Text. In Banking, Insurance, and Healthcare, Those Questions Come Before the Accuracy Gains. A Bad Movie Suggestion Costs A Viewer A Minute. The Same Pattern Applied To Credit, Claims, or Care Decisions Needs Lineage For Every Piece of Context The Model Reads.

An AI agent can give the right answer and still fail the test. Imagine a support agent that resolves a customer’s issue, but sends their personal data to an unauthorized tool along the way. If your evaluation only scores the response, that run could look successful. I’d want to know which identity made that call, what data the tool received, and whether we could have blocked the action or stopped the run. Those questions belong in the development process. That is part of what ADLC, the Agent Development Lifecycle, addresses. SDLC gives us established practices for building, testing, releasing, and maintaining software. MLOps adds versioned data, model evaluation, monitoring, and drift management. With agents, we also have to evaluate the tools and actions they choose during execution. Google Cloud’s Advent of Agents Season 3 opens with ADLC, then covers identity, security, deployment, observability, and operations throughout October. 𝗧𝗵𝗲 𝗳𝗶𝘃𝗲 𝗔𝗗𝗟𝗖 𝗽𝗵𝗮𝘀𝗲𝘀 -𝗦𝗰𝗼𝗽𝗲: Decide what the agent should do, who owns it, and where its authority ends. -𝗕𝘂𝗶𝗹𝗱: Develop with Agents CLI and ADK. Test its tool choices alongside its responses. -𝗦𝗰𝗮𝗹𝗲: Make deployment repeatable using infrastructure as code. -𝗚𝗼𝘃𝗲𝗿𝗻: Enforce identity, permissions, and policy throughout the lifecycle. -𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲: Use evaluations and execution traces to understand failures and improve the next version. Governance starts when you define the boundaries. It does not wait until step four. 𝗕𝗲𝘆𝗼𝗻𝗱 𝗔𝗗𝗟𝗖, 𝘁𝗵𝗲 𝗺𝗼𝗻𝘁𝗵 𝗰𝗼𝘃𝗲𝗿𝘀 • Identity and access: how an agent acts on someone’s behalf without gaining extra permissions. • Runtime protection: handling prompt injection, data leakage, and code execution risks. • Fleet management: knowing who owns each agent and which tools it can use. • Observability and containment: reconstructing what happened during a run and stopping unsafe behavior. • Day 2 operations: managing cost, catching regressions, and continuing to evaluate after deployment. My suggestion: pick one agent and use it throughout the month. After each relevant episode, add a control, an evaluation, or a trace you can inspect. Each episode stands on its own, so start with a problem you are already working on. And keep the support-agent example in mind: would your current evaluations catch that unauthorized tool call? The series runs October 1 to 31, with one question a day, a practitioner video, and runnable code. It is open to everyone, with no sign in or paywall. Make this your October learning plan: 31 days of practitioner videos and runnable code on building, testing, securing, and operating AI agents. Pick a problem you’re working on and build along:https://fandf.co/4iK9wX9 Thanks to Google Cloud for sponsoring this post.

If you want to actually understand LLMs and Generative AI (not just copy prompts), this Stanford lecture series is one of the cleanest step-by-step paths I’ve found. Here are the 9 sessions (watch in order): Transformer Fundamentals — https://lnkd.in/gaaRDexT Transformer Models + Practical Tricks — https://lnkd.in/gy4FUwNY Transformers → Large Language Models — https://lnkd.in/gsPiCrEU How LLMs Are Trained — https://lnkd.in/gvHJvgqP Tuning & Adaptation (fine-tuning, etc.) — https://lnkd.in/g6kgtPKR Reasoning in LLMs — https://lnkd.in/gAACSUG6 Agentic LLMs (tools, planning, workflows) — https://lnkd.in/gVm6js9z Evaluation: what “good” really means — https://lnkd.in/gJhbFQ4s Recap + What’s trending now — https://lnkd.in/g5JMNTsf Also, subscribe to this - https://appliedstack.ai/ Learn AI by doing - and https://brij.media/zero My suggestion: treat it like a mini-bootcamp, one lecture at a time, take notes, and pause to implement small experiments after each video. If you’re learning LLMs in 2026, this is a solid foundation.

A lot of LLM calls inside agent pipelines exist to answer one small question. Which queue gets this ticket. How risky is this request. Should this tool call run. TypeSafe AI released Jev in early access on September 15, and it is built for those calls. It is a decision model and generates no text. The animation shows the timing. Jev returns three answers in a single beat. The LLM builds a sentence word by word, then hands it to a parser before the app can use it. 𝗛𝗼𝘄 𝗝𝗲𝘃 𝘄𝗼𝗿𝗸𝘀 You send state (a ticket, a record, some JSON) plus typed questions. There are three question types: - Choice picks one option from a set you define - Score places the input on an ordered scale and can land between levels, like 1.4 - Noul returns the probability that a yes or no question is true Every question is answered in one pass, each with probabilities. The answer can only be a value you defined, so there is no text to parse or repair. 𝗪𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝗟𝗟𝗠 𝘀𝘁𝗶𝗹𝗹 𝗯𝗲𝗹𝗼𝗻𝗴𝘀 An LLM generates its reply token by token. You want that for reasoning, planning, writing, and code. For a routing decision it adds latency and cost, and your code still has to validate the text before acting on it. 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗴𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 Guardrails in an agent system are mostly decisions: should this action run, does this case need a human, is this output safe to send. A typed answer with a probability gives you a threshold you can tune, a value you can log for audit, and a clear rule for sending low confidence cases to a person. The policy itself stays in your code, where reviewers can read it. 𝗪𝗵𝗮𝘁 𝗜 𝘄𝗼𝘂𝗹𝗱 𝗰𝗵𝗲𝗰𝗸 𝗯𝗲𝗳𝗼𝗿𝗲 𝗮𝗱𝗼𝗽𝘁𝗶𝗻𝗴 𝗶𝘁 - TypeSafe has not published the model size, training data, or architecture - It is in early access, so pricing and limits may change - Vendor benchmarks tell you little about your own decisions. Pin a model version and test it against cases your team has already labeled If you were writing the rule for your team, what would decide whether a step in an agent gets a typed decision model or a full LLM?

If you want to actually understand LLMs and Generative AI (not just copy prompts), this Stanford lecture series is one of the cleanest step-by-step paths I’ve found. Here are the 9 sessions (watch in order): Transformer Fundamentals — https://lnkd.in/gaaRDexT Transformer Models + Practical Tricks — https://lnkd.in/gy4FUwNY Transformers → Large Language Models — https://lnkd.in/gsPiCrEU How LLMs Are Trained — https://lnkd.in/gvHJvgqP Tuning & Adaptation (fine-tuning, etc.) — https://lnkd.in/g6kgtPKR Reasoning in LLMs — https://lnkd.in/gAACSUG6 Agentic LLMs (tools, planning, workflows) — https://lnkd.in/gVm6js9z Evaluation: what “good” really means — https://lnkd.in/gJhbFQ4s Recap + What’s trending now — https://lnkd.in/g5JMNTsf Also, subscribe to this - https://appliedstack.ai/ Learn AI by doing - and https://brij.media/zero My suggestion: treat it like a mini-bootcamp, one lecture at a time, take notes, and pause to implement small experiments after each video. If you’re learning LLMs in 2026, this is a solid foundation.

This diagram shows every part of an AI agent at work. This animation follows one AI agent through a full run. The request is "Summarize Q3 churn, email the team." The agent goes around the plan, act, observe loop three times before the task is done. Each dot color shows what is moving: - Orange: the task and every command - Blue: results coming back from a tool - Pink: short term memory - Yellow: long term memory - Lime: trace events 𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝗶𝗻 𝗼𝗻𝗲 𝗿𝘂𝗻 Lap 1 runs a SQL query for the churn numbers. Lap 2 runs Python to build the chart. Lap 3 sends the email, and that step waits at the guardrails until a person approves it. At every PLAN step the LLM reads short term memory to see what it has already done. At every OBSERVE step the result goes to long term memory, and a trace event lands in the observability panel. 𝗧𝗵𝗲 𝗽𝗮𝗿𝘁𝘀 - Instructions set the goal and the limits before the first step - The LLM chooses the next step and hands the action to a tool - Memory carries context between steps and between runs - Tools act on real systems: a database, a code runtime, an inbox - Guardrails check every tool call before it runs - Traces record each step so someone can review the run later 𝗪𝗵𝗲𝗿𝗲 𝗜 𝗽𝘂𝘁 𝘁𝗵𝗲 𝗮𝗽𝗽𝗿𝗼𝘃𝗮𝗹 𝗹𝗶𝗻𝗲 The SQL query and the Python chart run on their own. They read data and produce a draft. The email reaches other people and cannot be pulled back once sent, so it waits for a person. I sort agent actions into three levels by what happens if they go wrong: reading data, changing internal records, and anything that leaves the organization. Each level gets a stricter check. Keeping that rule in the guardrail layer puts the policy in one place, where reviewers can read it, instead of spreading it across prompts. 𝗪𝗵𝘆 𝘁𝗿𝗮𝗰𝗲𝘀 𝗺𝗮𝘁𝘁𝗲𝗿 When an agent sends the wrong report, the first question is what it saw and why it picked that step. A trace that records the plan, each tool call, each result, and each approval answers that question. With only the final output, the team is left reconstructing the run by hand. Which agent actions do you think should always need a human sign off, however capable the models become?

This diagram shows every part of an AI agent at work. This animation follows one AI agent through a full run. The request is "Summarize Q3 churn, email the team." The agent goes around the plan, act, observe loop three times before the task is done. Each dot color shows what is moving: - Orange: the task and every command - Blue: results coming back from a tool - Pink: short term memory - Yellow: long term memory - Lime: trace events 𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝗶𝗻 𝗼𝗻𝗲 𝗿𝘂𝗻 Lap 1 runs a SQL query for the churn numbers. Lap 2 runs Python to build the chart. Lap 3 sends the email, and that step waits at the guardrails until a person approves it. At every PLAN step the LLM reads short term memory to see what it has already done. At every OBSERVE step the result goes to long term memory, and a trace event lands in the observability panel. 𝗧𝗵𝗲 𝗽𝗮𝗿𝘁𝘀 - Instructions set the goal and the limits before the first step - The LLM chooses the next step and hands the action to a tool - Memory carries context between steps and between runs - Tools act on real systems: a database, a code runtime, an inbox - Guardrails check every tool call before it runs - Traces record each step so someone can review the run later 𝗪𝗵𝗲𝗿𝗲 𝗜 𝗽𝘂𝘁 𝘁𝗵𝗲 𝗮𝗽𝗽𝗿𝗼𝘃𝗮𝗹 𝗹𝗶𝗻𝗲 The SQL query and the Python chart run on their own. They read data and produce a draft. The email reaches other people and cannot be pulled back once sent, so it waits for a person. I sort agent actions into three levels by what happens if they go wrong: reading data, changing internal records, and anything that leaves the organization. Each level gets a stricter check. Keeping that rule in the guardrail layer puts the policy in one place, where reviewers can read it, instead of spreading it across prompts. 𝗪𝗵𝘆 𝘁𝗿𝗮𝗰𝗲𝘀 𝗺𝗮𝘁𝘁𝗲𝗿 When an agent sends the wrong report, the first question is what it saw and why it picked that step. A trace that records the plan, each tool call, each result, and each approval answers that question. With only the final output, the team is left reconstructing the run by hand. Which agent actions do you think should always need a human sign off, however capable the models become?

Your AI agent is running. Should it be? An agent can complete every task but do you know if it’s doing what it was programmed to do or if it’s providing value?   Dataiku's Global AI Confessions Report: CIO Edition, 2026, based on a Harris Poll survey of 685 CIOs across eight countries, puts numbers behind that concern.  ● 90% said they were confident they had complete tracking of their AI agents. ● Yet 72% could not consistently confirm that those agents were delivering the business outcomes they were built for. That gap deserves attention from everyone building agents. A status dashboard can show that a workflow ran. The team also needs to know whether it shortened resolution times, reduced errors, or helped someone finish useful work. Before scaling an agent, I'd want six questions answered: 𝟭. 𝗪𝗵𝗼 𝗼𝘄𝗻𝘀 𝗶𝘁? A named person accountable for its purpose and performance. 𝟮. 𝗪𝗵𝗮𝘁 𝗰𝗮𝗻 𝗶𝘁 𝗱𝗼? Clear boundaries for data, tools, actions, and human approval. 𝟯. 𝗗𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? Evaluation on real tasks, including failure cases. 𝟰. 𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗶𝘁 𝗰𝗼𝘀𝘁? Usage and spend attributed to the agent and its use case. 𝟱. 𝗪𝗵𝗮𝘁 𝗱𝗶𝗱 𝗶𝘁 𝗶𝗺𝗽𝗿𝗼𝘃��? A business outcome measured against a baseline. 𝟲. 𝗪𝗵𝗼 𝗰𝗮𝗻 𝘀𝘁𝗼𝗽 𝗶𝘁? A clear path to pause, roll back, or retire it. These controls belong in a shared orchestration and governance layer, connected to the systems where teams build and run agents. Business teams should be able to build close to the problems they understand. Their agents still need visible ownership, operating boundaries, and measurable results. Pick one production agent, and ask yourself:  Can my team answer all six of these questions? Read Dataiku's Global AI Confessions Report: CIO Edition, 2026: https://lnkd.in/dkVi_Jnp   Thanks to  Dataiku for sponsoring this post.

Your AI agent is running. Should it be? An agent can complete every task but do you know if it’s doing what it was programmed to do or if it’s providing value?   Dataiku's Global AI Confessions Report: CIO Edition, 2026, based on a Harris Poll survey of 685 CIOs across eight countries, puts numbers behind that concern.  ● 90% said they were confident they had complete tracking of their AI agents. ● Yet 72% could not consistently confirm that those agents were delivering the business outcomes they were built for. That gap deserves attention from everyone building agents. A status dashboard can show that a workflow ran. The team also needs to know whether it shortened resolution times, reduced errors, or helped someone finish useful work. Before scaling an agent, I'd want six questions answered: 𝟭. 𝗪𝗵𝗼 𝗼𝘄𝗻𝘀 𝗶𝘁? A named person accountable for its purpose and performance. 𝟮. 𝗪𝗵𝗮𝘁 𝗰𝗮𝗻 𝗶𝘁 𝗱𝗼? Clear boundaries for data, tools, actions, and human approval. 𝟯. 𝗗𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? Evaluation on real tasks, including failure cases. 𝟰. 𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝗶𝘁 𝗰𝗼𝘀𝘁? Usage and spend attributed to the agent and its use case. 𝟱. 𝗪𝗵𝗮𝘁 𝗱𝗶𝗱 𝗶𝘁 𝗶𝗺𝗽𝗿𝗼𝘃𝗲? A business outcome measured against a baseline. 𝟲. 𝗪𝗵𝗼 𝗰𝗮𝗻 𝘀𝘁𝗼𝗽 𝗶𝘁? A clear path to pause, roll back, or retire it. These controls belong in a shared orchestration and governance layer, connected to the systems where teams build and run agents. Business teams should be able to build close to the problems they understand. Their agents still need visible ownership, operating boundaries, and measurable results. Pick one production agent, and ask yourself:  Can my team answer all six of these questions? Read Dataiku's Global AI Confessions Report: CIO Edition, 2026: https://lnkd.in/dkVi_Jnp   Thanks to  Dataiku for sponsoring this post.