By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Product SchoolZapier CEO on Rebuilding Products to Be Agent-First | Wade Foster
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
Great conversations run long. Their best ideas don't have to.
uListen is free on iPhone, iPad, Android, Chrome, and Claude.
See howCHAPTERS
Jump to publisher timestampsChapter summaries
Why Zapier keeps “staying alive” despite constant new model releases
Carlos frames the conversation around how each new OpenAI/Anthropic release seems to wipe out startups, yet Zapier continues to thrive. Wade attributes Zapier’s staying power to tuning out headline swings and focusing on consistently shipping what customers need over long time horizons.
- •Carlos argues new LLM releases repeatedly disrupt and kill startups, yet Zapier endures
- •Wade describes the current cycle of hype as “it’s so over / we’re so back” driven by headlines
- •Wade says Zapier’s approach is to listen to customers and keep shipping improvements
- •Wade suggests consistency over decades increases the odds of still being in business
Zapier’s foundation: fully remote, profitable, and built in the no-code era
Carlos highlights Zapier as an unusual company: fully remote and profitable. Wade confirms both and sets up how Zapier’s original product approach centered on making automation accessible to people who don’t write code.
- •Carlos calls out Zapier as fully remote and profitable
- •Wade confirms Zapier is remote and profitable
- •Wade describes Zapier’s roots as a pioneer of no-code automation
- •Zapier initially focused on UI abstractions on top of APIs to reduce coding toil
The wake-up call wasn’t ChatGPT—it was GPT-4 and the pace of improvement
Wade explains that while ChatGPT was impressive, GPT-4 was the moment Zapier internalized how fast model capabilities were about to accelerate. That realization drove two shifts: changing the product itself and changing how the company works.
- •Wade says ChatGPT impressed them but wasn’t the true wake-up moment
- •GPT-4’s quality, lower cost, and rapid release timing signaled an exponential improvement curve
- •Wade identifies two dimensions that needed to change: the product and Zapier’s internal ways of working
- •Zapier entered a period of experimentation across product features and internal workflows
Turning point: reasoning models “one-shotting” code and engineers not reading it
Wade describes a later inflection point when newer reasoning-focused models started reliably generating code in one shot. This changed Zapier’s conviction about the future: the product would evolve from abstracting code away to generating code on the user’s behalf.
- •Wade describes advanced reasoning models as the moment it felt “really here”
- •He says engineers increasingly prompted models to write code rather than reading it line-by-line
- •Tests/evals became a key way to check model-written code
- •Zapier’s product direction shifted from removing code complexity to having Zapier write the code for users
From no-code UI to “a code product users never see”: Zapier MCP and agent-built automations
Wade explains that Zapier’s fastest-growing area is Zapier MCP, which plugs into an agent so the agent can build automations for the user. He frames it as the original vision fulfilled: the best no-code experience is an agent that writes the code for you.
- •Wade says Zapier MCP is growing fastest because it can be installed into an agent of choice
- •The agent then builds automations and writes the underlying code
- •Wade calls Zapier a code-based product now, even if users aren’t technical
- •He compares it to having “the world’s best developer” next to you building integrations on request
- •Wade says the best no-code product was always an agent writing code for you
Why OpenAI/Anthropic haven’t “cracked it”: focus, speed of rewriting priors, and complementary strategy
Carlos asks why the big labs’ workflow offerings haven’t won. Wade argues focused teams beat generalist orgs on specific customer problems, and that enduring companies build alongside model progress—rapidly discarding outdated experiments and staying aligned to what their customers value most.
- •Wade claims focused teams on narrow problems can stay ahead of general capability builders
- •He says the labs’ core priority is making models better, while application attempts can be hit-or-miss
- •Wade argues entrepreneurs can extend models into specialized domains and novel user experiences
- •He emphasizes building alongside labs rather than competing “in front of them”
- •Wade highlights speed at “rewriting priors” when new model capabilities arrive
- •He says customer communities that value a company’s orientation can help it stay ahead in a dimension
Agent-first as the next computing platform: products agents can discover, buy, and use
Carlos raises the concept of being “agent-first” as agents increasingly act online. Wade describes agents as a new computing paradigm and defines agent-first products as tools that agents can sign up for, purchase, and operate end-to-end—without needing a human to mediate each step.
- •Wade says agents are becoming a new computing platform similar to OS, browsers, and mobile shifts
- •He frames an opportunity: identify what agents do well today and where they struggle
- •He suggests evaluating whether models/harnesses will solve gaps; if not, build tools agents can reach for
- •Wade defines agent-first: agents should be able to sign up, buy, and use the product like a human can
Keeping costs down: workflows over agents, deterministic execution, and model routing/ensembles
Carlos asks how to stay model- and tool-agnostic while avoiding surprise bills. Wade says many people overbuild agents when a deterministic workflow would be cheaper and more reliable, and then describes technical approaches like model routing and using multiple model providers (or no model) to optimize cost.
- •Wade says expensive usage is a real challenge right now
- •He argues people often build agents when they should have an agent build a workflow instead
- •He says Zaps run deterministically, making most workflows cheaper, more reliable, and more efficient
- •Wade references benchmarks showing workflows can be about 71% cheaper than re-reasoning each run
- •He describes using routers: frontier models delegating tasks to cheaper models when possible
- •He notes model routing can feel too complex for less technical users, creating product opportunity
- •He suggests ensembles across providers and even replacing models with deterministic code in cases
Breaking into enterprises: turning connector fear into a governance moat
Carlos asks how Zapier moved beyond indie hackers into large enterprises. Wade explains Zapier invested in enterprise features, but the crucial differentiator was governance around connectors—giving CIOs visibility and fine-grained control over data access so standardizing on Zapier feels safer than alternatives.
- •Wade says Zapier addressed enterprise “checkbox” needs like SSO, SCIM, and audit logging
- •He identifies the scariest category risk: connectors and where data can move
- •Zapier built governance systems on top of connectors to control read/write access and permissions
- •Wade says Zapier’s permissioning can be more fine-grained than typical OAuth defaults
- •He claims this flips connectors from a risk into a strength by increasing visibility and control
- •Wade describes a Fortune 500 company standardizing all connectors/automations on Zapier due to trust
Enterprise AI stalemate: “turn it all on” vs “shut it all down,” and the need for a third option
Carlos notes enterprise adoption often stops at basic assistant use because tool-to-tool automation feels risky. Wade empathizes with leaders stuck between enabling rapid innovation and taking on unacceptable risk, arguing the application layer can provide governance that unlocks a safer middle path.
- •Carlos says enterprises use assistants for simple tasks but avoid deeper automation due to fear
- •Wade says leaders feel trapped between enabling everything (powerful but risky) and banning it (kills innovation)
- •He warns that refusing innovation can drive AI-forward talent to leave
- •Wade notes governance decisions can get stuck in committees that talk instead of act
- •He argues application-layer governance can enable innovation while closing key risk vectors
- •Wade says Zapier aims to let many people build automations while shutting down risky behaviors
The discovery problem: agents that watch your work and recommend automations
Carlos asks how AI can help people discover which tools to use for a task. Wade describes a personal setup where agents review his activity logs across tools and provide weekly suggestions for how they could have helped, and he contrasts that with pre-AI recommendation systems that felt sterile and ineffective.
- •Carlos frames discovery as knowing which tools to use for a given job
- •Wade describes agents “watching” his work via logs from Slack, Gmail, Calendar, Docs, code, and data
- •He says the agent summarizes weekly and suggests ways it could have helped or what to automate
- •Wade says Zapier tried automation recommendations for years, but they were bland and didn’t work well
- •He says an agent-first approach produced better results even in a hacky personal setup
- •Wade notes the organizational challenge: giving agents broad visibility triggers governance/privacy concerns
- •He suggests prompting an agent to review recent work and propose automations as a starting point
Wade’s builder-CEO operating system: maker mode mornings and resisting meeting creep
Carlos shifts to Wade’s personal management style and time allocation. Wade explains he changed his schedule to protect daily maker time—especially to stay current on fast-moving AI—while battling the organizational tendency to add more meetings, including reducing frequent one-on-ones for his own role.
- •Wade says scaling led him into back-to-back meetings and low satisfaction about real output
- •He says many meetings delegated work talented teams could do without him
- •After having kids and seeing AI acceleration, he prioritized more maker mode to stay up to date
- •Wade set a rule: no meetings before 11am to preserve morning heads-down building/prompting time
- •He says meeting creep is an organizational tendency and requires constant pushback and subtraction
- •Wade mentions experimenting with meeting quotas and doing fewer one-on-ones with direct reports
- •He emphasizes protecting attention to avoid being outpaced by more AI-first competitors
Meetings as prompts, broad read-access for agents, and the self-improving company loop
Wade adds nuance: while many meetings are wasteful, debate-heavy discussions can generate valuable context that agents can turn into execution. He then explains that agents perform best with broad raw read access, cautious write permissions, and ongoing feedback loops that continuously surface signals for improvement—pointing toward the “self-improving company.”
- •Wade says meeting recorders plus models make some meetings a powerful prompt source
- •He describes debating a feature, then feeding the transcript to an agent to build what was discussed
- •Wade says agents improve with wide read access to raw data across systems (messier context helps)
- •He cautions that write access requires thoughtful controls and human-in-the-loop checks
- •Carlos references the “self-improving company,” and Wade agrees with the direction
- •Wade describes agents listening to signals like social complaints, customer calls, support tickets, and dashboards
- •He suggests agents can pursue goals continuously and improve systems via feedback loops
Publisher timestamps
- 0:00Trailer
- 0:53Intro: Carlos welcomes Wade Foster, CEO of Zapier
- 2:44Welcome, and why Zapier keeps staying alive
- 3:38Fully remote and profitable
- 3:48The repositioning: baking AI into the core product
- 4:33Why GPT-4, not ChatGPT, was the real wakeup call
- 5:55The turning point: models one-shotting code nobody reads
- 6:40Why Zapier is now a code product users never see
- 7:43Why OpenAI and Anthropic can’t crack it: focus beats general
- 10:09The two strategies: build against the models or alongside them
- 11:09What being agent-first actually means
- 12:50Keeping cost down: model-agnostic and tool-agnostic decisions
- 13:11Don’t build an agent, have the agent build a workflow (71% cheaper)
- 14:00Model routers and multi-model ensembles
- 16:03Getting into large enterprises, not just indie hackers
- 16:40Turning connector governance into the moat CIOs trust
- 18:54Why enterprises are stuck between turning AI on and shutting it down
- 21:10The discovery problem: which tools to use for a task
- 21:34Agents that watch you work and recommend automations
- 24:17Wade as a builder-CEO: structuring the day
- 24:29No meetings before 11am, and protecting maker mode
- 26:29Fighting the organizational urge to add meetings
- 29:38How debate meetings became a powerful prompt for agents
- 30:37Closing the loop: giving agents broad read access
- 31:45The self-improving company, and where this all heads
- 32:40Wrap-up