Recommendation · MoneyShubhankar Srivastava recommends protecting against prompt injection at the infrastructure layer using Browserbase safe mode to control which URLs the agent can visit and what actions it can take. You can also implement action permission modes like Codex or FlotCode — always allow, allow once, or allow any action — using Browserbase safe mode.
AI Engineer · Hill-Climbing Skills: Improve Agents Without Changing the Model — Shubhankar Srivastava, Browserbase
Recommendation · MoneyDakkak recommends using Modular's AI skills tooling to translate code from languages developers already know into Mojo or Max, and warns that rewriting from scratch wastes time when porting tools exist. He also suggests testing the tools on a small module before a full migration.
Advice from the speaker; effectiveness of the tooling is not independently verified.
AI Engineer · Taming the AI Hardware Cambrian Explosion — Abdul Dakkak, Modular
Recommendation · MoneyJonathan says he has been pushing his team to use Jam MCP as the standard way to capture UI/UX details and acceptance criteria from non-technical people, including recordings that capture console logs and network requests that can be fed into an AI harness.
AI Engineer · Build with Perception Agents — Emile Baizel & Shruti Arora, Amazon AGI Lab
Recommendation · MoneyDan recommends measuring accuracy including hallucination rate — 3% in his example — determined partly automatically and partly via manual annotation.
AI Engineer · https://www.youtube.com/watch?v=PwCSYy-oGsM
Recommendation · MoneyUse the MCP search engine API first to find target pages, then call the data pipeline API to pull structured results — a pattern illustrated by a client building a social media listener who needed follower counts across Instagram, TikTok, and LinkedIn.
stated
AI Engineer · https://www.youtube.com/watch?v=IK8cC4FScps
Recommendation · MoneyWhen building a resume-matching model, source jobs data that includes which companies are posting, how many jobs are available, salary, and location — and ensure data flows to you consistently because performance depends on continuous delivery.
stated
AI Engineer · https://www.youtube.com/watch?v=IK8cC4FScps
Recommendation · MoneyThor says the easiest way to try the Gemini Live API is through Google AI Studio: go to ai.studio/live (or ai.dev/live) to talk to the model and explore its capabilities. In AI Studio you can select Gemini 3.1 Flash Live, choose a different voice (e.g., Puck), and enable grounding with Google Search. The Live Translate model can also be tried there.
AI Engineer · Can Your Agent Hear You Now? Building Live Voice Agents with Gemini — Thor Schaeff
RecommendationRun custom harness judges in CI on agent-proposed PRs, experiment with them, and optimize from there.
endorsed
AI Engineer · Building Self-Learning Loops for Your Agent — Fuad Ali, Arize AI
RecommendationHarness-as-a-judge can connect to external systems and call them to verify whether the user was allowed to do the action and whether the trace reflects the access they had.
endorsed
AI Engineer · Building Self-Learning Loops for Your Agent — Fuad Ali, Arize AI
RecommendationA price-correctness rubric passes when every returned product's price is within the requested min/max range and fails when any result exceeds the budget. The rubric can go into the code, check filters work correctly, and catch the issue on every subsequent run — not just a pre-coded test case.
endorsed
AI Engineer · Building Self-Learning Loops for Your Agent — Fuad Ali, Arize AI
RecommendationEnable Signal to run on top of your agents to get clustering analysis of issues. Configure it to trigger alerts off a specific eval threshold (e.g. an eval that must stay below a 1% trigger rate), a latency spike, or a cost spike. It takes a few minutes to get started once traces flow in.
endorsed
AI Engineer · Building Self-Learning Loops for Your Agent — Fuad Ali, Arize AI
RecommendationWith Harness-as-a-judge, write the rubric as a plain-language sentence; it reads the actual trace directly and runs in a sandbox against real traces, instead of writing column mapping and eval template boilerplate.
endorsed
AI Engineer · Building Self-Learning Loops for Your Agent — Fuad Ali, Arize AI
RecommendationWith Arize skills, you can ask your coding agent (e.g. Claude Code) to go read your traces from within the repo, seeing what the agent is doing from trace data and the actual code at the same time. Example prompt: 'Hey, check my traces for any errors or anything I can fix. Cross-reference with the code for the agent in this repo and use the agent skills.'
endorsed
AI Engineer · Building Self-Learning Loops for Your Agent — Fuad Ali, Arize AI
Recommendation · MoneyTo use Data Designer for RL, skip the golden solution generation step and end up with a dataset full of prompts. Use NeMo Gym, which is essentially an RL reward function tool: you write a function that gives you a reward, then use any RL training framework to execute in real time with those tools and provide a reward.
AI Engineer · Teaching Agents to Search with NVIDIA Data Designer — Dhruv Nathawani, NVIDIA
Recommendation · MoneyAlways use the cheapest correct column type when building data generation pipelines. Using an LLM to create a sampler is a complete waste of resources — write samplers in Python instead — but using an LLM to generate text data makes a lot of sense.
AI Engineer · Teaching Agents to Search with NVIDIA Data Designer — Dhruv Nathawani, NVIDIA
Recommendation · MoneyUse pass-fail checks on actual output plus LLM-as-judge for impact analysis tasks when evaluating agentic systems.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Recommendation · MoneyTo build a code-based graph analysis agent, you can leverage call graphs at SCIP to understand who's calling this and who references this. It can read the comments inside and outside of the code, and fold in prose from outside of the code, including GitHub issues, linear issues, and tickets.
endorsed
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Recommendation · MoneyYoni Michael recommends mining gold data from the graph but cross-referencing it back to the source code so the gold isn't cheating — i.e., not polluting gold context with graph output.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Recommendation · MoneyYoni Michael recommends hooking the code graph in as an MCP to the tools you're already using — e.g., having the code graph in addition to Claude Code and asking Claude Code to call the graph — and then writing skills so the agent uses the graph properly. To engineer an agent to use these things in actual coding, promote the prompt-engineering practices into skills and give those to your coding agent. He notes an engineer receiving such a PR for the first time can get a half-decent first pass from being more deterministically guided, and that the deterministic part is helpful because you can incrementally diff over what's changed and analyze it deterministically, getting a fast conclusion about what's happening as opposed to waiting for the agent to read everything.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Recommendation · MoneyThe host recommends using a classification layer in front of an agent to check for prompt injection. For such a layer, you can either use an LLM-based layer, or alternatively a semantic search layer (e.g., vector search via Quadrant or LanceDB), or an in-memory process.
AI Engineer · How Secrets Leak Through AI Agents (and How to Stop It) — Venice AI