35,595 results · Most relevant
Price a one-to-many membership with your team at roughly five times the one-on-one price (around $8,000–$10,000) or get rid of it. Pricing a team membership at $150/month when one-on-one is only twice the price breaks the value ladder and hurts scaling. Include base-level support in the standard offer and charge separately for one-on-one access.
Alex recommends building a proper CS playbook to onboard and train team answers — reviewing all past responses, making suggestions, and letting the owner just approve.
Alex recommends capturing 30-40 posts plus green screens, then taking the winners and running them as ads to the wait list or webinar, noting this means 20-30 posts a day per platform.
When a business has a very high close rate (85-90%), increase price because there's probably a lot of room there. Alex Hormozi believes the financing adoption rate could be pushed to 80-90% with the right payment framing.
endorsed
Robert Brennan recommends getting a separate computer — a virtual machine in the cloud or a Mac Mini — and putting all your agents on it as a third sandboxing strategy, which gives you a long-lived environment with persistence. He notes the caveat that co-located agents can still interfere with each other.
Robert Brennan recommends running coding agents inside Docker containers on your laptop as a first sandboxing strategy: spin up a brand-new container for every new conversation with the agent, which is how OpenHands originally launched. Because the container is ephemeral, the agent can run wild — if it runs 'rm -rf /' and deletes everything, you just kill the container and start again.
The guest recommends that if the community exists elsewhere, build a subreddit and build a community. A rough benchmark for deciding whether to do so is about 20,000 branded searches a month — which is quite high — though the benchmark may be a little lower if your listenership or customers have a high average customer value.
moderate
The host recommends selling globally by default from day one rather than expanding market by market, saying the entire world is your market. The old SaaS pattern of focusing on your home market first and expanding later just doesn't work anymore and is far too slow compared to top AI companies. Likewise, the old playbook of building for a long time before getting your first customer is too slow for the current market reality.
Nnenna Ndukwe demonstrates opening a pull request and having Kodo review the PR directly in GitHub. In the PR review and remediation gate, Kodo's findings are inspected and fixes and deferrals are documented. You can then review the issues Kodo found and resolve them automatically with an agent skill, and rerun the verification command after fixing the coding changes found in PR review — looping rather than stopping at the first local pass.
Nnenna Ndukwe maintains planning templates that apply to any repo, defining the expected plan structure regardless of the codebase. The planning template includes a summary, problem statement, constraints, scope, behavior scenarios, verification gates, failure and recovery rules, a commit plan, and skill routing to pre-plan which skills the coding agent will kick off during implementation. The build execution template documents the starting state (what branch, current local verification process, expected commands) and includes a definition of done or completion note. Document everything in the plan including deferred items and changed decisions.
Endorsed; no stated uncertainty.
Run threat modeling scans on every code or AI application change pushed (e.g., pull requests). Set up webhooks that trigger a scan whenever something changes, run a scan whenever infrastructure-as-code is changed, and monitor Terraform files created by agents — running a scan the moment they change. Integrate threat modeling early when designing AI-agentic workflows.
The speaker recommends using automated tools to check that any MCP server you install is free of malware before running it for the first time. He admits he does not personally check all the code of every MCP server he installs, but he limits a lot what he installs.
Provide guidance to the agent at the start of optimization to steer it from the beginning. Point the WeCo agent to a markdown file containing all your information and context so the agent has the best chance at finding success. For MiniPG optimization, this can include hard constraints and promising ideas like focusing on compression, different quantization methods, and weight sharing across model blocks.
Shubhankar Srivastava recommends: if you want to ship skills in production, bake them into your agent loop. Build systems where as soon as you run into any production failure, you run an auto research loop to see what went wrong, heal it, and create a pull request that goes directly to a human engineer who fixes it, and the skill gets updated automatically. Alternatively, run a sample run every two or three hours to see if the skill is working perfectly — there are different ways of having cron jobs check skill accuracy. On Anthropic managed agents, run a cron job using the skill every day at 9 a.m., configured to send a notification on failure, which triggers a healing loop to handle websites changing or anti-bots changing — baking in checks and balances around changing websites and antiB
A proposed approach: replicate applications (like world models) so you have better control and guardrails, then let the agent do free exploration in the environment about 100 times so it becomes familiar with the environment. After exploration, the goal is finding the first successful trajectory; once found, it can be successful all the time. For an initial run on a program the agent has never seen (e.g., a legacy payroll process), the initial prompt has to be very specific, e.g., go to an employee, go to payroll, check this box.
proposed approach; 'a couple of months' out for the current product
Coach Kyle recommends starting conversations with strangers by asking a simple, low-pressure question about something in the environment, calling it the 'obvious question' technique. He says it works in any setting and is one of the easiest first steps toward becoming more social.
Coach Kyle recommends committing to giving out compliments and greetings to at least three people per day. He frames comfortable greetings and compliments as the foundational level-one step of his social roadmap, and says taking this one step consistently can open many doors and build social connections.
Dhruv Nathawani recommends giving the agent a single tool (Tavily Search) in the system prompt, adding usage guidelines so the agent does not over-call the tool and rack up large costs, and including an in-prompt example so the model learns how to call the search tool via in-context learning.
Omar urges the audience to send the Epstein cover sheet to ChatGPT and try to get results out of it, to see how accurate (inaccurate) frontier models are on complex layouts.
The agent hallucinates when a tool fails instead of disclosing the failure, so prompt-based changes are needed in addition to evals for tool failure. Duggal recommends building an eval on top of it to detect when the tool is failing AND making prompt-based changes so the agent doesn't hallucinate. He also suggests instructing the agent to build the tool-success/silent-failure eval and fix numeric groundedness so responses are always grounded to an actual reference point instead of made-up numbers.
Choose sources → to get a library of your own and a daily brief.