35,684 results · Most relevant
Pass the session ID back in as resume so the next turn remembers everything the last turn saw. Put the two turns together in one trace, wrapped in trace agent to get one parent span encompassing all the data. Without wrapping you would get two sibling spans, which is fine but slightly harder to read.
Francesco recommends being hand-wavy with reviews prior to functional tests passing, but never pushing anything into production that hasn't passed all your tests.
Francesco recommends that when creating a harness, you build gates not only for validating things but also for correcting the course and making sure you can do so safely.
Approach AI adoption from a highly impact-based point of view rather than 'we should use AI for this because AI can something, something, something'.
endorsed advice
A structured template for in-season lifting: lower body day one pairs a knee-dominant squat pattern with a jump, plus a Nordic or Razor Curl supersetted with a 3D lateral lunge for the groin, and mobility/core work. Upper body day one pairs a horizontal push with a dynamic horizontal push, an upper back movement with a vertical pull, and optional grip work. Lower body day two uses a split squat pattern or trap bar deadlift with a split squat jump, hip-dominant posterior chain work (hip thrust or RDL) with groin work, foot and calf work, plus core and mobility. Upper body day two uses a vertical push (landmine press) supersetted with a rotational mid-back row and a horizontal row for upper back/scap work, plus biceps, triceps, and arm care in coordination with the training staff. Exercise
Consequential creator advice requires independent verification before Today can promote it.
Grubbs passes along advice to young women entering men's pro sports: 'make it about the role, not about the gender, like just dominate the position that you're in.' When asked how she'd get respect from male athletes, she answers 'yeah' and just goes and does it. She also coaches nervous interviewees: 'these are your talking points, go in, dominate.'
To find the most likely failure in language models, use perplexity: search for scenarios with low log perplexity and high toxicity (the top left corner of the search space). If the most likely failure (toxicity above some threshold) is really likely, study it and put in place mitigations.
AI systems should be built with zero-trust architectures so that infiltrating one system does not allow infiltrating everything else.
You should continue the validation effort after deployment, not just before; this is sometimes called runtime monitoring or runtime assurance. You want to validate what's going into your decision system, i.e., the sensor inputs; work with Airbus was done on integrity monitoring of vision-based systems. You also want to monitor what is happening in the internals of the system, which has led to work on mechanistic analysis for an aircraft system. And you also want to model and monitor the outputs of the system.
Continuously monitor AI benchmarks for distribution shift, just as you would monitor AI systems, and use quantitative methods to identify shift in language — because definitions of harms drift fast as new generations introduce new slang while benchmarks move slowly.
Split the dataset because people are lazy to filter — get at least a golden trace dataset, which may be much smaller, and fine-tune that with a much heavier weightage.
Efficiency in the memory layer comes from the number of entities that you create. Keep the memory layer simple: don't create too many abstractions or too many components. Experiment first with only short-term memory and long-term memory — just two structures.
creator's recommendation
Nacho Martinez advises deciding your own memory layer architecture based on the use case rather than adopting a fixed design.
Nacho Martinez recommends keeping the context window as small as possible, since the smaller it is, the better the model solves the problem.
Implement human-in-the-loop feedback in a flow by giving Claude a custom tool called 'request human feedback' that Claude calls whenever it needs input.
The speakers recommend using Claude Code (UltraCode) when sitting at your laptop typing into the terminal interactively, and Cloud Managed Agents (Flows) when you have a repeatable thing that should happen on a regular basis and want to integrate against an API — treating UltraCode as the single one-off instance and Flows as the layer to reuse and run agents on a trigger schedule.
The most common way to give this tool to a customer is to build a product interface (website or app) as a layer on top of your managed agent; a user action like dragging a card from one column to the next on a Kanban board might trigger a Claude Managed Agent session in the background via the API.
Prompt a flow with 'Go out and scan across all these services and tell me if we're healthy' plus 'don't touch prod', or during an outage: 'We are having an outage right now, please go look at these 50 different services and tell me the root cause of the outage, but don't touch production.'
Use a markdown file that links to other markdown files to gradually reveal context to subagents and save tokens; when a subagent only needs context relevant to its slice, have it read the linked markdown details relevant to that sub-agent.
To create an agent, pass in a multi-agent config to enable multi-agent; pass in pre-configured sub-agents in the agent config if you have specialists important to your research process — e.g., agents specializing in Yahoo Finance, news sources, or SEC Edgar; to get flows up and running, send a first event to the lead agent saying 'run a workflow' — that's all you need.
Choose sources → to get a library of your own and a daily brief.