Recommendation · MoneyPienaar recommends generating many different theories simultaneously to give the agent breadth and diversity of ideas, calling this Cleric's primary technique. The rationale: agents are excellent at relative ranking but terrible at absolute scoring of their own theories, so multiple candidate ideas beat asking for one scored answer.
AI Engineer · Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, Cleric
Recommendation · MoneyPienaar recommends computing a baseline for production — extremely important in large environments — because most prod environments are in permanent disarray and chaos: engineers know which error logs to ignore and which metrics are unused, but an agent doesn't. Concretely, cluster logs and give the agent statistics on which logs recur daily and are common, not just raw logs; the same approach applies to metrics, object state, and many other things.
AI Engineer · Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, Cleric
Recommendation · MoneyPienaar recommends using a topology or build dependency graph of the production environment. When the agent claims a causal connection between services, output all connections that actually exist and have another agent critique the claim — an effective way to introduce doubt and push the agent to go further if the claim doesn't map to a real connection.
AI Engineer · Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, Cleric
Recommendation · MoneyPienaar recommends that when sub-agents hand context to each other, they build a case with evidence rather than just giving the answer. This mirrors what SREs do — 'I looked at this metric, and it was at this point in time' — hard evidence rather than 'prod is up.' If evidence isn't propagated, it loses fidelity and causes hallucination and overconfidence.
AI Engineer · Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, Cleric
Recommendation · MoneyTo know whether techniques actually work beyond vibes, Pienaar recommends looking at verification and outcome in addition to diagnosis. The production incident response flow with agents is detection, diagnosis, and fix, and most folks only use agents for diagnosis. Cleric triangulates across many different signals what happened when a fix is applied — monitoring whether the engineer actually merges code that maps to the diagnosis — and checks for a return to normal in service metrics, logs, and baseline state. Muting an alert also shows a return to normal, which is why multiple signals must be checked. This outcome machinery exists to calibrate evaluation.
AI Engineer · Your AI Agent Is Confidently Wrong About Production — Willem Pienaar, Cleric
Recommendation · MoneyUsing web search, an agent can look at the repository itself, find breaking changes from the repo's changelog, and explain the migration in depth. For a code review agent, a rule might be: when the diff bumps a dependency or version, look at the upstream source to verify it is actually true, and then ground the review in what you find.
Stated as a suggested rule ('might look like'), not a mandate.
AI Engineer · Your Coding Agent Is 6 Months Out of Date — Jakub Hojsan, Exa
Recommendation · MoneyTo enable easy multi-agent systems, developers must solve four key layers: Transport (ordered/real-time delivery, retrying), Continuity (persistency, hydration, runtime binding), Conversation Abstraction (rooms, channels, participants), and Governance (identity, audit).
AI Engineer · Your Agents Are in Solitary Confinement: Why MCP & A2A Aren't Enough — Vlad Luzin, Band
Recommendation · MoneyBuild 'narrow startups' targeting very small user bases with high willingness to pay ($200-$300/month) to achieve high run rates without mass market scale.
high
The a16z Show · The Personal Agent Race Is Here | Anish Acharya & David Pawlan
Recommendation · MoneyRequire explicit user permission before an agent takes direct actions that impact the user (e.g., switching insurance), but allow proactive action for tasks requiring no user-side action (e.g., drafting emails, getting credits).
high
Contested by Jed (Harrison Stall) on AI Engineer
The a16z Show · The Personal Agent Race Is Here | Anish Acharya & David Pawlan
Recommendation · MoneyIn high-autonomy environments, employees must develop strong judgment ('spidey sense') to determine when to pause, seek feedback, or perform spot checks before pushing work live, balancing speed with quality.
high
The a16z Show · Building a Team at AI Speed | Harvey’s Maggie Landers
Recommendation · MoneyTara Seshan recommends identifying the eigenquestion — the specific most important thing to test. Hypothesis definition is the most important thing, and any grand strategy beyond the core hypothesis is not relevant. Be as pointed as possible about your core hypothesis.
Lenny's Podcast · AI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)
Recommendation · MoneyTara Seshan distinguishes two types of writing: 'writing as thinking' (outlining, writing prose, cutting, editing, iterating to get ideas in line) and 'writing as reporting' (status updates, summaries). She happily automates reporting with AI but never automates thinking. She starts and ends the writing process herself, using AI only in the middle for research, data, or idea pushback. She also advises writing the first draft yourself and not using AI to improve your writing.
Lenny's Podcast · AI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)
Recommendation · MoneyFor a bootstrapper, a feedback loop longer than 12 months is dangerous because you risk wasting years of your life. The opportunity cost of spending 12-18 months on a long sales cycle is significant. If you have a long sales cycle, look for a vertical where something closes sooner to optimize learning speed. Evaluate whether you can learn faster than a long enterprise sales cycle by finding smaller customers or an MVP approach. If you pursue a long sales cycle, either raise funding, have personal runway, or do it nights and weekends. If you haven't closed a deal, it is very hard to raise funding unless you have a great network or a prior exit.
Startups For the Rest of Us · Episode 847 | What Second Time Founders Do Differently, Pricing AI Agents, and More Listener Questions (Rob Solo)
Creator perspective · MoneyFrank explains that each line of a pipeline has a runner, and there are many ways to run these functions: code, LLM, statistics, person. He recommends using deterministic code for simple conditional logic (e.g., if the price-to-earnings ratio is less than 15) because code is fast, cheap, and reliable. Use statistics and classical machine learning to model more complex relationships than pure deterministic code. Use LLMs and AI agents for the most complex relationships — e.g., summarization is the state of the art for LLMs. It's probably totally impossible to build a bunch of if statements that can summarize better than a modern LLM.
hedged
AI Engineer · AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue
Principle · MoneyNikhil Harithas recommends standardizing on a compute gateway abstraction that agents call into, rather than binding to a specific compute provider, as the abstraction needed for this pattern.
AI Engineer · Your AI Agents Shouldn't Care Where They Run - Nikhil Harithas, Factory
Principle · MoneyNikhil Harithas says an important idea in Open Compute is that it is a project to manage computers and not sessions. Tying tooling to a session concept makes the software vendor-specific, which makes customers uncomfortable hosting it.
hedged
AI Engineer · Your AI Agents Shouldn't Care Where They Run - Nikhil Harithas, Factory
Principle · MoneyNikhil Harithas endorses suspend, snapshot, and fork as the standard pattern for managing sandboxes. Forking is especially important: as fleets and swarms of agents gain steam, one session delegating to another with sub-agents is a fractal idea, and being able to save a snapshot, fork it, and hand off a subtask to another agent will be really important. Snapshotting is also going to be really important for agent infrastructure generally.
AI Engineer · Your AI Agents Shouldn't Care Where They Run - Nikhil Harithas, Factory
Creator perspective · MoneyThe revenue velocity model tracks four metrics: number of opportunities coming in, ACV (opportunity size), win rate, and sales cycle length. The host recommends using it in the first 30 days to determine what actually needs your time versus what is painful but won't move the needle. Katie also ties each metric to specific enablement skills so metric declines point to skill gaps.
30 Minutes to President's Club | No-Nonsense Sales · #622: Sales Enablement Playbook: How To Win Your First 90 Days
Creator perspective · MoneyThe speaker summarizes with five questions to ask when building AI features: are people using it or just saying they will use it; do you read real outputs or just look at the averages; does it still pay off after factoring in all the costs; is a win on the best model something you can afford; and are you clear legally on which markets you can ship in. The question is no longer just whether the feature won the A/B test.
hedged
AI Engineer · 78% Wanted This AI Feature. Almost Nobody Used It - Supriya Subudhi, Bumble
Principle · MoneyZach Lloyd says the factory should have a strict definition — workflows, skills, MCPs, agent definitions, automations, benchmarks, webhooks, and where agents run — stored in a code repo. Without this, you lose the ability to see the full factory definition at any point in time, version control its parts, test different versions, and let coding agents automate updates to it.
AI Engineer · How to Build a Software Factory That Improves Itself - Zach Lloyd, Warp