35,848 results · Most relevant
Set data invalidation rules such as age-based invalidation — for example, data more than a month old should never be considered. (The speaker framed the one-month threshold as an example, not a universal rule.)
example threshold, not a universal rule
When pulling in a customer's full set of tickets, summarize them and store the summary in your structured stored context. Storing derived summaries prevents having to re-look up and re-summarize everything for common questions in the future.
The bottleneck was never reasoning depth, it was just tool use — improving the reliability of that is where the model was actually failing and improving. Once the model masters the fundamentals, it can compose those skills.
Charles Dickens argues that specialized workflows don't actually need generalist models — you don't need polymaths. The reflex when improving agents may be to scale, given scaling law results, but specialized workflows don't need generalists. As an analogy: for a tax audit you would call the specialist that knows the forms, the tools, and the rules, not Terence Tao.
Small models can successfully be trained with RL when you have quality data to represent the target outcomes and tasks. This approach can result in better deployment economics for specialized tasks and domains. This result has been repeated in other areas including health care, law, and insurance.
A product at 80% reliability feels untrustworthy to customers because they still have to monitor the agent and do manual work when it fails — ending up with more work, not less. Around 92% reliability, the message changes: customers feel they can trust the product to handle their tasks and completely hand over the workflow.
Memory systems need consolidation so duplicates can be resolved and conflicts handled. They need to evolve over time, otherwise they become stale, irrelevant, and useless. Memories need an element of temporality — when created, when accessed, and how long they are valid for; a memory may only be valid for a certain time window. And memory systems need regular feedback and evaluation: if you're not measuring use and sending feedback into the memory layer, it will all fall over.
Secrets and personal information need to be redacted before they hit the memory layer. Rights need to be scoped — to the agent, the workflow, the team, the user — and governance controls are especially important in multi-tenant environments. Memory systems also need appropriate admin approvals, review, and the ability to suppress stale or poison data.
Viren Baraiya's key framing: your harness is the application and vice versa. An agent alone doesn't deliver the whole premise; a harness that controls the agent's execution delivers on the premise. He recommends building a harness that controls the execution of the agent.
Viren Baraiya explains that a harness runs more than just an agent loop; it is a combination of deterministic and non-deterministic parts of the application. Non-determinism is delivered by an LLM in terms of reasoning and thinking process, and you do not want non-determinism sipping into the deterministic part of the application.
Define a loss function to tell the network when it makes a mistake: the closer the prediction is to the actual value, the lower the loss; the farther apart, the more error incurred. When training on many examples, use the empirical loss function, which is the mean of all individual loss functions across the dataset. The goal of training is to identify the set of weights W* that minimizes this average empirical loss — the whole goal of the class is learning how to identify W star, training all the weights in a network such that the output loss is as small as it can possibly be.
None stated.
Regularization is a method introduced into training to discourage complex models and encourage simpler models to be learned. Regularization is critical for models to generalize past training data. A model should be sized to balance generalization capability with capturing overall trends in the data — aim for a model size in the middle: small enough to maintain generalization capabilities but large enough to capture the overall trends.
None stated.
Gelsinger argues every fundamental element of virtualization and management — workload, network, storage, security — needs to be recreated for the AI agent computing hierarchy, creating many new company opportunities. He endorses building VMs designed for agents rather than humans, since agent usability as the primary design constraint changes form factor, marketing, allowable complexity, and required startup speed (agents are far less patient than humans). Vendors should optimize so the agent — not the human — makes the purchasing decision.
Em Shreve argues that simply handing people AI tools does not produce good quality learning or work; people need real time to learn and experiment, and an environment for learning should be the star of any education program.
A high-level context graph view helps answer questions about engineering architecture that could not be answered in the context of a single repo or microservice; once the graph existed, the team could ask previously impossible questions.
The centralized validation phase is on the critical path of getting changes into production, so faster is always better.
Patel argues the goal of adopting AI should not be efficiency but benefiting from the original insights AI generates to solve previously unsolvable problems. If you try to go efficiency first, you'll never get familiar or get good.
Will Lyon argues reasoning memory belongs alongside short-term and long-term memory as a first-class citizen, and that agents should store their thinking and the facts — including how the agent decided — not just the data.
The ideal flow for internal tools is that the creator builds the app, deploys it, shares it, and everyone gets access without engineer involvement. WorkOS wanted everyone at the company to be able to build and ship an application without asking infra for provisioning or getting security approval for standard apps, and wanted instant distribution: ship an app, link someone in Slack, they click and access it without red tape.
endorsed
Matt Gibiec argues that an agent does not need a nervous system or sight to be efficient and productive; what it needs is actual context of your environment. Context is king — it is the nervous system you can provide to your agents, and it acts as connective tissue driving all the processes you automate.
Choose sources → to get a library of your own and a daily brief.