33,905 results · Most relevant
The first step to improving acceleration is to assess it objectively with metrics: maximum force output, maximum power output, and propulsive power output on the ground.
Matt Brockman recommends giving sandboxes a really long time to live so they can run as long as the task needs (e.g. a one-hour command timeout), but once the task finishes, set a very short lifetime (about one minute) so the sandbox shuts itself down and frees resources for another sandbox.
A common sandbox pattern is to install base dependencies (like GitHub Actions caching), then fork the sandbox and build on top of the replica. When starting with a vanilla sandbox, keep adding to it and fork it so others get a replica; eventually bring it all together. This also enables forking a data science workload into multiple sandboxes to run experiments in parallel, and handing off in-progress work to others by sharing a sandbox template.
Matt Brockman's core lifecycle policy: give sandboxes a really long time to live while a task runs (e.g., one-hour command timeout), but once the sandbox finishes its task, give it a very short timeout (one minute) so it shuts itself down and frees resources for someone else's sandbox. This applies to unpredictable workloads like reading large files or video processing. A participant also confirmed you can pause or shut down a finished sandbox to resume it later if the user returns.
Matt Brockman recommends spinning up a new sandbox specifically for a project so you can break it without breaking your initial stuff; if something breaks, just restart the sandbox — the reload button resets it to its original snapshot state.
At RunLayer, agents do the real work; similar agent runs are grouped together and a skill is distilled from them — when a run succeeds and contains something interesting (multiple tool calls), they take the trace and ask a frontier LLM to distill a skill from it. Distilled skills pass through a series of checks, including whether they contain prompt injection or PII, before being accepted. Once a skill passes its checks, it lands in the skill library, and an MCP gateway serves approved skills to each and every client and agent within the company. Distilled skills are included in all future agent runs so each run benefits from prior discoveries.
When rolling out AI strategy across a company, govern skill distribution from one centralized place (the MCP server) rather than scattered local installs, giving one place to manage what knowledge is shared, with policy enforcement and role-based serving.
Take the ten life categories and rank them one to ten for yourself; then identify which priority is off — where your stated ranking does not match your actual effort. The speaker's one tactical tip: do things that check multiple boxes, like lifting with coworkers, which covers recreation, career growth, and physical health, letting you block more time with no extra time required. You won't get this for everything, but occasionally you get lucky.
The lifting example is the speaker's personal anecdote.
Ali recommends having really good metrics in place as a platform team, driving enough standards across the organization to get telemetry on how fast code ships to production, and measuring engineer throughput, quality/rework, and engineer sentiment. He stresses having developer productivity metrics as a baseline before adding AI so you can measure, iterate, and adjust, thinking in terms of the entire lifecycle and entire workflows rather than one tool in one part, offloading entire workflows to agents that autonomously act and deliver outcomes, paying attention to developers' time and keeping them in flow, and automating away grunt work since agents are perfect for it. He notes developer sentiment is equally important, if not the most important thing.
endorsed
Tariq recommends being super clear about what you are measuring, avoiding token maxing in favor of real value, being very deliberate about the process you put in place, orchestrating around a guide-verify-solve model, constructing tooling and systems deliberately, getting the team to trust the agents, and setting up lighthouse teams to highlight best-practice ways of working with agents. He adds that role modeling alone does not get other teams to adopt; they must also understand why they are using the agents.
endorsed
Dor Sasson recommends that with AI products, enforcement should happen synchronously in the hot path before balances are updated, with reconciliation and settlement done asynchronously after the fact — like an ATM, where the check of whether you're allowed to draw happens before you get the cash.
Speaker's endorsed recommendation
Ben recommends applying the Git-like branching workflow to database schema changes: branch the schema, make changes in an isolated environment, and merge back to production via a deploy request with no downtime — plus the ability to one-click revert a deployed change back to the old database state without data loss.
Evals are the starting point and ending point for any agent architecture. Think about evals in five levels: positive reinforcement, negative reinforcement, tracing the path to the solution, scoring like an engineer, and calibration/confidence. Score the AI like an engineer by comparing how it performed the investigation against how the best engineer would perform it, and evaluate how the AI calibrates itself and gives out its confidence level.
Define the specific guardrails AI needs to operate under, based on your team's access levels, restrictions, and industry. Define specific conditions for write access — read, write, and under what conditions write is allowed.
The root cause must always be based on a causal chain of evidence showing the exact steps that led to the issue. If a causal chain of evidence cannot be established, report the information with a low confidence level and point the user in a different direction rather than fabricating a root cause.
Marc Klingen describes a changelog writer agent workflow: the agent updates public assets (documentation, changelog, social posts) based on what has been shipped; it suggests changes to documentation via a merge PR; there is a code review step as a GitHub PR where someone can approve and merge or submit change requests; the changelog writer acts on change requests to update the draft until it is finally released; and approvals and requests-for-changes are fed into AI observability and evals to serve as the basis for auto-improvement of the agent.
Dr. Masud Husain describes factors that may confer cognitive resilience to Alzheimer's pathology: a routine of physical exercise, keeping blood pressure from getting too high, checking for and avoiding pre-diabetes or diabetes, keeping cholesterol under control, not drinking too much alcohol, and not smoking. Psychological factors — having a sense of purpose, maintaining a network of people you're socially connected with as you age, and remaining curious and open to new ideas — turn out to be almost as important as the physical health measures. He notes we are just scratching the surface of understanding how purpose, social connection, and curiosity affect the biology of cognitive resilience.
explicit
Dr. Husain recommends that to overcome someone's activation energy barrier, you can either make the effort less or change the incentive. Break a task down into subcomponents so the activation level doesn't seem so formidable. Increase the reward or change it to something meaningful for the person — the reward becomes the pleasure they get from the new framing.
Dr. Husain advises that while intrinsic attention training may lack strong evidence, you can cut out things that are distracting to rescue degraded attention. Concretely: turn off notifications (which many people don't do), close your email while working on something else, and ensure there's no distracting noise. Modifying your environment this way caps the drain on your attention.
Dr. Husain describes his 'deathbed decision' test: when making decisions about what to do, choose only things you could look back on from your deathbed and think were valuable and worth doing. If an option would not be something you'd be proud of or value in any way, just say no to it.
Choose sources → to get a library of your own and a daily brief.