Recommendation · MoneyJohn Stecher recommends using agents to help understand existing human workflows, then writing code to optimize for straight-through processing and injecting agent intelligence only where humans previously made decisions. He also advises stepping back to identify work agents can do better, to offload intellectual cycles from investment, finance, and engineering professionals so they can do higher-value work.
AI Engineer · Why Waiting for the Next AI Model Can Pay Off - John Stecher, Blackstone & Scott Worland, Norm AI
Recommendation · MoneyContinuously test the workflow at every part of the workflow, designing the test to that specific part and measuring success in that element. For a 100-step workflow, each step must be evaluated for what goes into it, what comes out of it, and the expected outcome. Testing only the final output misses intermediate failures, like an image edit that changed the background but added unwanted handles. Assuming input and output to be the outcome of anything in the workflow leads to untested failure points.
AI Engineer · If Your AI Scores 100%, Your Eval Is Wrong - Ramakrishnan Lokanathan, eslams.com
Recommendation · MoneyCharlie Holtz advises creating 'slot-free zones' — parts of the code base or app that require really strict human review. He warns that if you're not careful about slot-free zones, your code base can get in a really tricky spot, and notes Conductor has had to rewrite their whole app a couple of times because of this. As a concrete example, any change to the migrations file in Conductor's CI requires a human to review it. He also notes this careful-in-places approach is unusual: most people assume top builders are pure token maxers ripping through 30,000-line PRs.
AI Engineer · Orchestras, Not Factories: How the Fastest Builders Work — Charlie Holtz, Conductor
Recommendation · MoneyDaksh Gupta recommends optimizing the prompt as the lowest-hanging fruit for improving a review agent. Start with a very simple prompt like 'review this PR and output the comments in GitHub's API schema format', then orient the prompt around two themes: look for specific bug types, and skip the nitpicks. Do internal data science on existing outages and incidents to figure out what types of bugs your codebase has the highest propensity to produce, create a catalog of those bug types, and have your prompt look for them. Also expressly tell the review agent what a nitpick is and not to look for them. He cautions that changing a small thing in the prompt has a weirdly big impact on the number of comments the agent writes, which is quite scary.
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Recommendation · MoneyDaksh Gupta recommends model inversion: use a different model to review code than the one that wrote it — if Claude Code wrote the code, use a GPT model to review it, and vice versa. The types of bugs produced by Claude models and Codex models are quite different from each other (Claude models are something like three times more likely to produce one-off errors than GPT models, while GPT models are significantly more likely to produce auth bypasses than Claude models), and these differences are artifacts of post-training, as far as he understands. Using the same model to review code that the same model wrote results in relatively fewer bugs caught, because the model fails in the same ways during review as during generation. Greptile found a 10% higher catch rate if you flip models, and a 1
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Recommendation · MoneyDaksh Gupta recommends building a dataset of historical PRs that caused outages and incidents to evaluate your review agent: catalog pull requests that have caused outages and incidents in the past if your company nicely catalogs when things go wrong, run every version of your review agent through that dataset and measure with what frequency it catches those bugs, and keep the dataset updating in real time because the topology and ontology of bugs changes over time in a changing world. This allows you to tune the prompt more confidently. He notes you can get decently far evaluating with your internal data, especially for your particular company.
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Recommendation · MoneyIdo Salomon recommends the software-factory pattern: give an agent a general sense of what you want, and it spawns an orchestrator, breaks the work down, and runs it all in a local container. Because everything runs in isolation, he says you don't need to babysit the work.
AI Engineer · We're the bottleneck, but we don't have to be — Ido Salomon, MCP Apps
Recommendation · MoneyKevin Hou endorses using sidecar processes to let the model listen to the outside world and set up its own triggers, hooking them to SMS messages, webhooks, cronjobs, or GitHub PRs. He also endorses using scheduled task crons in Antigravity for time-based automation.
AI Engineer · Get Out of the Model's Way — Kevin Hou, Google Antigravity
Recommendation · MoneyNoyan recommends scoping agents to a task they're good at rather than having one agent do 10 different things: narrow scoping makes agents easy to measure and easy to lifecycle-manage, and keeping things measurable is really important when designing agents. Every agent is versioned, so you can push a new version to add capabilities in small increments. Making a new agent that is a union of existing agents' functionality often ends up with a lot of tech debt and works poorly, and if you bloat an agent some functionality gets better while some gets worse and you can't easily tweak it. Going forward, the firm encourages people to leverage their agent gateway rather than building one agent with many functions.
hedged
AI Engineer · How a Trading Firm Ships New AI Models the Same Day - Noyan Tokgozoglu, Optiver
Recommendation · MoneyRoman Ugarte recommends using GrokBot as an infovore that consumes huge quantities of information, removing it from your cognitive load and coming to you with what's important. The V1 implementation: GrokBot sits on top of Slack and email, you tell it your role and what you care about, which cases warrant a direct notification, and which should just go into a daily roundup you read every day. He also gives bots a complete fire hose of information, having hooked his up to every mention of GrokBot on X plus internal context, and describes an always-on chief of staff entity that preserves your focus on what matters.
Commercial context: The creator presents this as their own offering and directs viewers to get it.
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyRoman Ugarte recommends setting up GrokBot to consume large quantities of information and remove that cognitive load, acting as an always-on chief of staff that preserves focus on important things. He suggests connecting it to Slack and email, telling it your role and what you care about so it notifies you only in important cases, including non-urgent items in a daily roundup instead of pinging directly, and only allowing it to page you for super urgent matters if you trust it has no false positives.
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyAamir Shakir recommends giving the agent the choice between grep, normal search, or agentic search depending on query complexity.
AI Engineer · Better Search Took This AI Agent From 36% to 98% Accuracy - Aamir Shakir, Mixedbread
Recommendation · MoneyAamir Shakir recommends using agentic search for complex qualitative queries, which runs multiple queries to support multi-hop reasoning.
AI Engineer · Better Search Took This AI Agent From 36% to 98% Accuracy - Aamir Shakir, Mixedbread
Recommendation · MoneyKumar recommends keeping tools like AlphaSense or Yipit but bringing them into an internal platform that learns from analysts' habits and is programmed specifically for each individual analyst and their model pattern. The goal is to bring all the tools different analysts use into one platform and up-level the existing spend, increasing the value of data subscriptions by continuously learning from all the different types of data and model updates, leveraging the MCPs and APIs these companies offer. He adds that as AI adoption increases, especially at smaller fundamental shops, having one unified layer doing many different tasks across analysts will be really valuable.
AI Engineer · Proving Alpha: How a Hedge Fund Measures AI's ROI - Hari Kumar, Contour Asset Management
Recommendation · MoneyArcolano recommends getting to 30+ runs of using skills per engineer per week, which yields a 30% bump in PR throughput holding other factors constant — a bigger effect than model choice. Skill use includes cursor rules, harness customization, and tricks of the trade that must be learned.
AI Engineer · Better AI Models Won't Fix Your Engineering Team — Nicholas Arcolano, Jellyfish
Recommendation · MoneyMeasure what the finished job cost and how long it took, including retrieval, tool calls, and retries. If a person must fix the report before use, count that human work in the cost of producing the acceptable report. Include all the failed attempts in the cost of producing the acceptable report. Measure the cost of reports that users can actually use.
AI Engineer · Is a Faster Model Actually Faster? Test the Whole Job — Morgan Suo, MiniMax
Recommendation · MoneyChris Dreyer obsesses about personal injury law — he has books over his shoulder on personal injury lawyer marketing and personal injury lawyer intake. He hosts a Personal Injury Lawyer Mastermind podcast with 450-something episodes, runs a newsletter with several thousand subscribers, and is launching a print magazine and conference (PEMCON). He advises building an ecosystem of niche content to understand who your customer is, what their problems are, and how you can provide more value.
Ahrefs · How I Make $32M A Year Dominating This SEO Niche
Recommendation · MoneyGary Vaynerchuk's prescription (advice to Amber, generalized to the audience): make organic social media content over and over and over again, in as many different ways as possible, across Facebook, Instagram, TikTok, YouTube Shorts, and LinkedIn, using different thumbnails and different hooks — with complete insane psycho discipline.
Consequential creator advice requires independent verification before Today can promote it.
GaryVee · How I'd Grow on Social Media If I Started Today
Recommendation · MoneyGaryVee identifies hyper-local Facebook ads targeted to a 10-mile radius of a physical location as the biggest opportunity for a local business to drive foot traffic and increase average receipt size.
GaryVee · What to Post When You Run a Small Business
Recommendation · MoneyThe guest's formula for winning on Snapchat Spotlight is to first win organically, then post-produce the content with performance elements that match the most native elements of what is currently working on Snap.
GaryVee · Everything You Know About Social Media Is Changing