Good morning, AI enthusiasts!
When an agent does something wrong, the model usually gets the blame. But many failures start earlier, in what the system around the model shows it, keeps, and allows.
An agent can miss the right tool because it never appeared in its context. It can lose a rule when a long conversation gets summarized. It can make a change that should have needed approval. And a run can pass its final checks while the trace shows weak work underneath.
This week’s articles look at how to design and test those parts of the system. You’ll learn:
Plus: a simple check for catching unsupported claims, my workflow for preparing talks with Claude, and a community game that teaches agent engineering in a 3D universe.
Let’s get into it!
This week in What’s AI, I’m sharing the process I use to prepare conference talks with Claude. I start with rough notes, turn them into a clear story, and only move to slides once the structure works. From there, I use voice notes to review the deck, rehearse the full talk, cut long sections, and repeat until the timing and transitions are right.
The article covers the presentation skill I use, how I structure the story before touching the slides, how I review and rehearse, and the small things I’ve learned about presenting clearly and confidently. By the end, you should have a clear process you can use to improve your own next presentation, whether you’re a developer, an AI engineer, or not technical at all. It is simply the way I prepare, along with the tips that have helped me give better talks regardless of the topic.
Read the full article here or watch the video on YouTube. This week, while putting together a list of ways to reduce hallucinations, we used a simple check from Anthropic’s guide, and it proved quite helpful.
After the model writes an answer from your documents, ask it to review every factual claim and find a word-for-word quote from the source that supports it. If it cannot find one, remove the claim and leave [] in its place.
The empty brackets make review much faster. You can immediately see which parts of the answer were unsupported instead of checking every sentence from scratch. If the same section repeatedly comes back empty, the problem may be upstream; for example, the source material does not contain the information the model is being asked to use.
This does not eliminate hallucinations, and high-stakes outputs still need human review. But it gives you a cheap way to catch unsupported claims before they reach someone else.
The same check becomes harder to manage once an application is producing thousands of answers. At that point, you also need to track which failures are recurring, which source or prompt caused them, and whether a change actually reduced the error rate across your evaluation set.
That is the kind of work we get into in AI Engineering for Production, where we walk through how to use checks like this and turn them into evaluation workflows you can keep running as the system changes.
You can join the waitlist to hear when the book is released. — Louis-François Bouchard, Towards AI Co-founder & Head of Community
Nikodimapps has built Way of the Agent, a game where you learn to build AI agents step by step. The learning path is designed as a 3D universe you fly through. It includes 218 lessons and a small project in every module. It also has a mini-game where you can catch prompt injections and data leaks. This could be a fun way to learn new topics. Check it out here and support a fellow community member. If you have any questions or feedback, share them in the thread!
The Learn AI Together Discord community is flooding with collaboration opportunities. If you are excited to dive into applied AI, want a study partner, or even want to find a partner for your passion project, join the collaboration channel! Keep an eye on this section, too — we share cool opportunities every week!
-
Ddanon912 is looking to partner with a developer to develop AI solutions, contribute to challenging projects, exchange ideas, and create opportunities for long-term professional growth. If this sounds like the right path for you, connect with them in the thread!
-
Signorbuco’s company trains and develops LLMs, and they are currently looking for a collaborator with experience in SFT, LoRA, QLoRA, AdamW, and PyTorch. If this is your focus, reach out to them in the thread!
-
Geronimo.stilton36 is building a security runtime for AI agents and is currently looking for developers to help test, break, and challenge the design. If you would like to contribute, contact them in the thread!
Meme shared by hudsong0
The Agentic Data Development Lifecycle (ADDLC) by Savva Lukhnev
AI agents can make data-pipeline development much faster, but failures still carry the same cost once those pipelines reach production. This article introduces the Agentic Data Development Lifecycle, a vendor-neutral framework for deciding what agents can change, what controls must enforce, and where humans still approve decisions. It covers three actors, three feedback loops, seven phases, and four approval gates, with autonomy based on reversibility and blast radius. It also explains why Build and Run should stay separate, how master-data changes should be handled, and reports early results of roughly five times shorter lead times and eight times more integrations.
This article follows an operations agent that rolled back a deployment with kubectl, only for Argo CD to restore the broken version because the agent never found the tools referenced in the runbook. It covers six ways to improve tool selection, including scoping tools by agent, choosing tools in code for specific steps, deferring tool , improving names and descriptions, testing search with realistic user phrasing, and handling empty results. It also includes trace checks for confirming which tools actually reached the model.
A useful agent harness can start with one clearly defined workflow instead of a general-purpose architecture. This article begins with a research job card and builds a roughly 250-line Claude harness around its requirements. It covers source tracking, token budgets that include compaction costs, write restrictions, and evaluations for the completed workflow. It also shows why trace review matters even when the final outputs pass, and compares Sonnet with Haiku, which passed the same evaluation suite at about one-sixth of the cost.
- Make Your AI Agent Remember What It Was Told When the Chat Gets Long by Satsawat Natakarnkitkul (Net)
Long-running agents can lose important rules when frameworks drop old messages or replace them with a summary. This article opens with an inbox agent that was told not to delete anything until the user confirmed. Before compaction, it deleted mail without approval in none of 100 sessions; after compaction, it did so in 30. It explains that LangChain, Google ADK, Pydantic AI, the OpenAI Agents SDK, and Microsoft Agent Framework keep the full history until the model provider rejects the request as too long, which is usually when teams turn compaction on. It then covers five patterns for what the model should read on every call: repeating standing rules, keeping task facts in application state, auditing summaries, reducing tool-output size, and separating user memory from task state. It also includes framework hooks and trace checks for testing whether instructions survive long runs.
World models need to represent how the physical world changes without spending all their capacity generating every future pixel. This article looks at JEPA-based approaches that predict future states in latent space instead. It covers V-JEPA 2.1, LeWorldModel, H-JEPA, D-JEPA, and LeHami, including how they handle physical representations, training, planning, decision-making, and structured dynamics. It also covers compute and latency constraints for running these models on robots, discusses Z Robotics’ Trudy model, and compares prediction accuracy with decision quality as evaluation targets for physical AI.
If you are interested in publishing with Towards AI, check our guidelines and sign up. We will publish your work to our network if it meets our editorial policies and standards. LAI #152: Is Your Agent Failing Outside the Model? was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.