35,595 results · Most relevant
When building a production agent, tighten the access so you control which domains it can reach out to or which specific MCP servers it can use. For a demo, the speakers gave the agent unrestricted access, and in this demo specifically access to a few websites, mainly the SEC, plus Python packages so it can reach the SEC.
Lotte Verheyden advises taking enough time to make sure traces are set up well, because a bad trace means you won't catch everything during monitoring, leading to worse data sets and worse decisions.
Patterson recommends making sure you have some way of seeing everything an agent does before deploying at scale. For developers running Claude on YOLO mode to crank out PRs, he stresses ensuring the agent isn't doing things that would get you in trouble — because someone will be held accountable when an agent does something unexpected.
Pedro Lopez says skills shipped with a CLI are a must: without one, an agent has no way of knowing the CLI exists, and skills also help the agent use the tool effectively. Airbyte's skill surfaces with a description whenever connectors are asked about.
Teams should be held accountable for the work they are producing. Just knowing that someone else will review your work forces engineers to pay more attention to finer details and be more meticulous, especially if there are accountability metrics that can be made visible for each chunk of work product. Rewards should be centered around producing value, elegant well-factored pragmatic system design, and clarity of tests and evaluation harnesses.
Pay attention to everything she says during the interaction. Her statements reveal her situation (only here a couple days, on vacation, party plans), which shapes how you close.
The host recommends developing an unshakable core emotion that invites other people to follow you, presenting it as the third core charisma trait and noting that Trump and Mamdani stayed with their core emotion through attacks.
Dru Knox recommends making the process a bunch of small lifts rather than one massive snake eating the moose all at once — find one improvement at a time, take a workflow, turn it into a skill, ship it to everyone, and get improvement loops going.
endorsed
Dru Knox recommends setting up standards for agentic code review as well as more targeted checks using a suite of tools for code review.
endorsed
Dru Knox recommends setting up something that looks at agent output quality on a consistent basis and brings improvements and learnings back to the rest of the codebase, including updating the relevant skill or playbook when the agent makes a mistake so it doesn't repeat it.
endorsed
Dru Knox says the first layer of harness engineering work is getting your control plane right, and that per Tessl this must be done first, in order.
endorsed
Dru Knox recommends finding a way to identify repeated tasks and automate them as quickly and easily as possible, because if automation is hard, people won't do it. This includes mining PRs and issues for repeated tasks and setting them up as skills.
endorsed
Dru Knox recommends moving workflows onto surfaces where all information is saved and publicly available, since valuable harness engineering data is otherwise hidden in local agent logs or in people's heads.
endorsed
Sam recommends using an AI assistant that gives critical feedback and tells you when your idea isn't as good as you think, rather than one that acts as a cheerleader. He suggests a test: ask the assistant a question and notice if it always responds with 'great question' regardless of topic, which signals a yes-man assistant.
Anish Acharya argues that company building is increasingly becoming a series of creating loops — from per-person loops to company-wide loops. He recommends designing your company this way and using humans to help the loop escape local maxima and land at the base of the next hill.
Anish Acharya advises that when building AI products, you should think about the 10x or 1000x version of the idea rather than defaulting to a minimal MVP. He frames this as a habit of asking 'What's the 10x, the 1000x version of this?' to force bigger thinking about AI product opportunities.
Coach Kyle advises opening conversations in everyday settings like a deli line without doing a direct open or displaying intent, instead using low-pressure openers such as asking a question, giving a compliment, making an observation, or giving a greeting.
Uncle Bob built software for telephone-line technicians who climbed poles; his boss insisted he get on the truck and watch users in the field. After seeing the software used on the pole, he could think of about 30 improvements to make it easier for that guy to use, and he said it totally changed the way he thinks about building software for people.
Adler advises that when facing a crossroads between a pure-art path and a commercial path, the commercial route has lower stakes and a much greater ability to try and fail; he chose commercial design over the gallery art world and is glad, since a viable art-market career would have constrained him to one style forever.
Use datasets and experiments to iterate on your agent and actually measure whether it's getting better. Evals give you a number that you can track, compare, and act on.
Choose sources → to get a library of your own and a daily brief.