Recommendation · MoneyThe guest recommends enterprises use deterministic controls — sandboxes and human-in-the-loop approval gates — to feel safe deploying computer-use agents, even when the model itself is highly aligned.
The a16z Show · OpenAI: When AI Agents Use Computers Faster Than You
Recommendation · MoneyMasaba Gupta advises making everything on brand and never going off brand, keeping the storytelling and language the same across all brand touch points — she calls it the brand's 'body language'. She stresses that the story created in the head office must permeate all layers of the brand: retail, e-commerce, social media, communication, and all touch points, and that when sitting down with each team, whether creative or retail, she delivers a singular message.
Figma · Config India 2026: What makes the House of Masaba feel like Masaba
Recommendation · MoneyZach Lloyd recommends defining your software factory in code, with a strict definition covering the workflows, the skills, the MCPs, and where these run. Version-controlling the factory definition lets you test different versions of your factory and lets coding agents update your factory. He suggests having an actual code repo that sits behind the factory, where you can see a definition of the entire factory at any given point in time, including everything that goes into automating development: agent definitions, automations, benchmarks, where agents run, the skills, and the webhooks — all stored in a repo that allows for automation.
AI Engineer · How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp
Recommendation · MoneyZach Lloyd recommends setting up a self-improvement loop where agents look at failure modes, aggregate them, look for patterns, and come up with ideas to improve the factory definition. Agents can propose fixes to failure modes via a PR to the factory definition. One example fix is having agents re-read matching repository skills before implementing. He frames closing the loop — agents identifying problems and then fixing them — as a core concept for improving development.
AI Engineer · How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp
Recommendation · MoneyZach Lloyd recommends storing the entire factory definition in a code repo that allows for automation, including agent definitions, automations, benchmarks, where agents run, skills, and webhooks. He argues that defining the factory in code and version-controlling it enables testing different versions of the factory and lets coding agents update it — which is what enables the factory to improve itself.
AI Engineer · How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp
Recommendation · MoneyStandardize evaluation infrastructure (reusable frameworks, metrics, judges calibrated with human feedback); automate feedback loops (continuous trace collection, recurring failure detection, root cause analysis); scale automated optimization (prompt optimization, agent reflection, repeatable candidate testing); validation and trust (holdout evaluations, regression testing); and democratize RSI with reusable components and self-service info.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Recommendation · MoneyStart with prompt updates as the easiest and most straightforward way to iterate; then optimize the structural context of the AI system; then expand scope to automatically design the entire workflow (meta information design); then optimize the harness code itself; and finally, the most complex step, optimize the optimizer code itself.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Method · MoneyThe RSI process proceeds in ordered steps: (1) run the agent and generate samples — for the career agent, given a query, generate output and log the process in traces (e.g., LangSmith traces); the agent trajectories and traces generated are the material used for optimization. (2) Run the traces through the evaluation engine (e.g., the LLM judge) to get judge output telling where the system is good and where it is not, to guide future iteration. (3) Agent reflection (also called root causing): based on the judge output, understand the root cause of the bad output. (4) Based on the identified root cause, generate an edits proposal — currently the edits are prompts, but in future they can be new workflow designs, harness code, or optimizer code. (5) Validate the edits, because some edits can
Method, not a verified protocol.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Recommendation · MoneyDustin advises limiting SaaS demos to no more than three features — never going over three — and picking out the features you're going to demo in advance. When you have a large feature set and limited demo time (30-40 minutes), he recommends scoping down to drive home key points. He cautions against letting prospects drive the feature selection, because you end up so in the weeds with your technical baby that it derails the demo.
MicroConf · SaaS Sales Demos: Why Getting a No Is Better Than a Yes
Recommendation · MoneyDustin Perrier says you need to have a next action from every demo, and that you need to send a follow-up email the same day as the demo, with a little area you can personalize. He also recommends putting demo prospects into a drip campaign, because they may be busy after the demo and forget how wonderful your product is.
MicroConf · SaaS Sales Demos: Why Getting a No Is Better Than a Yes
Recommendation · MoneyDustin Perrier argues that problem statements are horrible because they're not emotional, and that a problem statement won't draw prospects into your conversation and presentation. Instead, he recommends crafting a story that walks prospects through: this is the problem I experienced, this is why I wrote this software, this is what it solved and what it hasn't solved.
MicroConf · SaaS Sales Demos: Why Getting a No Is Better Than a Yes
Recommendation · MoneyDustin Perrier argues that roughly 20% of a demo should linger on the intro slide of the deck, using that time for discovery instead of powering through the deck. He advises not proceeding until you have verifiable data that the prospect is engaged, and suggests using the intro slide as the conversation point where you take notes on prospect concerns and what you need to show to address them.
MicroConf · SaaS Sales Demos: Why Getting a No Is Better Than a Yes
Recommendation · MoneyBe agnostic of the model and harness: use third-party providers and on-prem models, all going through the inference gateway.
AI Engineer · How a Trading Firm Ships New AI Models the Same Day - Noyan Tokgozoglu, Optiver
Recommendation · MoneyIf your domain produces expert feedback, compiler error messages, profiler messages, or tool call error messages, return that as side information to the optimizer.
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
Recommendation · MoneyRoman recommends creating a very small team that goes off into a cave for about a month with the sole objective of building an amazing project, completely isolated from the rest of the company (private Slack channels, a separate part of the office), starting completely from scratch rather than adding to an existing product, and making many micro-decisions every day to move quickly.
Lenny's Podcast · How we built Grok Bot in a month | Roman Ugarte (SpaceXAI)
Recommendation · MoneyIrina advises against a single vague metric like 'agent quality', which hides interesting failures, and recommends a balanced scorecard instead. This is the measurement backbone for the scenario-based loop.
AI Engineer · Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati
RecommendationElie Bakouch recommends that a real benchmark should run multiple seeds with all models in the same condition, and that validating a record should require passing a statistical threshold so it is not just seed optimization.
AI Engineer · We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect
RecommendationElie Bakouch recommends designing benchmark tracks at three access levels: one with no external access to measure pure model capability from weights alone, one with only archive papers to measure literature-grounded research capability, and one with full access including the latest human records to measure best-possible performance.
AI Engineer · We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect
Recommendation · MoneyJemima Abu shares the lightweight version of the accessibility skill she uses: it tells agents to use semantic HTML only, reach for ARIA when needed, and ensure interactive elements work by keyboard alone.
AI Engineer · Teach Your AI Agent to Build Products Everyone Can Use - Jemima Abu, CAIS
Recommendation · MoneyJemima Abu recommends auditing your existing code base for accessibility faults before introducing your agent, because an agent introduced into a code base with poorly written code will take that bad code as the standard.
AI Engineer · Teach Your AI Agent to Build Products Everyone Can Use - Jemima Abu, CAIS