What is one specific practice or design choice that has been critical to successfully integrating agentic AI into complex infrastructure or workflows?

The answers emphasize building trust, safety, and oversight—through human vetting, error recovery, guardrails, and incremental deployment—while splitting between redesigning workflows for agents and layering in accountability or technical controls.

16 answers

Answers

  • Karim Lakhani

    Founding Chair HBS AI Institute · Harvard University

    Counterargument-First Prompt

    Karim Lakhani points to a counterargument-first system prompt, citing Kai-Fu Lee's Claude prompts, so the AI opens by challenging the user's premise and flags weak claims rather than agreeing, since it's easy to fall into uncritical acceptance.

    The answer starts at 7:41.

    Silicon Valley Girl: AI, Tech and Career Growth · Listen on Apple Podcasts
  • Brendan McCord

    Founder & Chair · Cosmos Institute

    Liability with Insurers

    Brendan McCord proposes liability as one of three layers, arguing that labs and insurers should be accountable and put money behind their safety judgments, with insurers pricing that risk.

    The answer starts at 4:56.

    MTS · Listen on Apple Podcasts
  • Rekha Shenoy

    Chief Executive Officer · BackBox

    Offline Non‑Traffic‑Impact Capabilities

    Rekha Shenoy says building trust in agentic systems requires focusing on offline non-traffic-impact capabilities that solve a real problem, with humans vetting remediation tasks before automating them over time.

    The Everything Feed - All Packet Pushers Pods · The Everything Feed - All Packet Pushers Pods · Listen at 18:23
  • Molly Graham

    Host, WorkLife Podcast · Ted Conferences

    Coach It Like an Intern

    Molly Graham says AI needs the same coaching and training a human does, comparing it to an intern, and argues people should iterate on and edit its output rather than copy-paste it.

    The answer starts at 35:51.

    Lenny's Podcast: Product | Growth | Career · Listen on Apple Podcasts
  • Dwarkesh Patel

    Host · Dwarkesh Podcast

    Dedicated Agent Environments

    Dwarkesh Patel says his team lets Grok Bot run its multi-step animation workflow on its own cloud computer, where all the tools it needs are installed, so it can handle the process end to end.

    Dwarkesh Patel · Watch on YouTube
  • Reid Hoffman

    Co-Founder, Executive Board Chair · Manas Ai

    Take Baby Steps

    Reid Hoffman says the practice is to take baby steps, starting by proving the AI can write the right email before making it more agentic and proactive.

    Reid Hoffman · Watch on YouTube
  • Anish Acharya

    General Partner · Andreessen Horowitz

    Internal Training Program

    Anish Acharya points to companies like Kavak running an internal training program, a six-week course teaching everyone including mechanics to use new tools, which he says shows more people embrace the technology than is often discussed.

    The answer starts at 8:27.

    a16z Podcast · Listen on Apple Podcasts
  • David Vellante

    Co-founder, co-CEO SiliconANGLE, theCUBE, theCUBE Research (formerly Wikibon) · Siliconangle & Thecube

    Preserve an Immutable Record

    David Vellante says companies should preserve an immutable record connecting original intent, identities, actions and resulting changes, because a log of technical events alone isn't enough for the company to reconstruct its intent and decisions when something goes wrong.

    SiliconANGLE theCUBE · Watch on YouTube
  • Tara Seshan

    Member of Product Staff · Openai

    Agents Working Locally

    Tara Seshan says agents working locally are really convenient because they have access to all the data on your machine, unlike cloud agents that require extensive infrastructure and system access to be effective.

    Lenny's Podcast: Product | Growth | Career · Lenny's Podcast: Product | Growth | Career · Listen at 19:08
  • Angie Jones

    VP Engineering, AI Tools & Enablement · Block

    Delegate Well‑Defined Tasks

    Angie Jones says she delegates tasks that are already well defined with success criteria attached, adding that she would not give agents relational work like messaging people on her behalf.

    Quality Remarks - The Podcast · Quality Remarks - The Podcast · Listen at 1:13:57
  • Angela Jiang

    Head of Product, Claude Platform · Anthropic

    Error Recovery Mechanisms

    Angela Jiang says building infrastructure that lets agents recover from their errors is needed, since models now run longer on their own instead of being steered in one direction.

    Kleiner Perkins · Watch on YouTube
  • Katelyn Lesse

    Head of API Engineering · Anthropic

    Agent-First Workflow Redesign

    Katelyn Lesse says the better approach is to break an ambitious project into basic tasks and reinvent the process agent-first, because conforming the agent to what a human was doing doesn't work and the agent ends up reinventing the workflow anyway.

    Kleiner Perkins · Watch on YouTube
  • Sam Altman

    Chief Executive Officer · OpenAI

    Robust Accident Reporting

    Sam Altman points to the FAA's extremely robust accident reporting as an approach that works well for new technology, and says his team will do good accident recording, study failures, and publish clear postmortems as they deploy models.

    David Senra · Watch on YouTube
  • Ryan Lopopolo

    Member of Technical Staff · OpenAI

    Guardrails

    Ryan Lopopolo says it is humans' role to give agents the tools, context, guardrails, coaching, and trust needed to fulfill the job they want them to do.

    Agentic AI Weekly · Agentic AI Weekly · Read
  • Alex Salazar

    CEO · Arcade.Dev

    Separate Authorizer from Executor

    Alex Salazar says the thing that takes an action should never be the thing that authorizes it, arguing the control has to sit outside the agent, which should act on your behalf rather than as you.

    The Control Layer with Amer Altaf · The Control Layer with Amer Altaf · Listen at 1:05:39
  • Aditya Khandelwal

    Member of Technical Staff · Amazon

    Self‑Healing System

    Aditya Khandelwal says you have to close the loop and build a self-healing system, because slop is inevitable and will seep in, so you need a pipeline to detect and remove it.

    AI Engineer · Watch on YouTube

More Generative AI questions

View all →