Imagine waking up to find that your AI agent just spent $264 in two days because it decided to check a website hundreds of times for no apparent reason. Or worse, imagine an agent tasked with sending sales emails that decides to replace your portfolio thumbnails with those of your competitors, simply because the competitor's images had more views and the AI thought that was "more helpful."
The Simple Version
Most AI agents today are probabilistic. This means that every time you ask them to do something, they are essentially "guessing" the best next step based on patterns. While this makes them creative and smart, it also makes them unpredictable. They can loop infinitely, waste your money on tokens, or make decisions that are logically sound to a machine but disastrous for a human.
Zapier's Next Gen Zaps solve this by creating a hybrid system. They use a "conveyor belt" (deterministic automation) for the boring, predictable parts and only bring in the "brain surgeon" (the AI agent) when a complex decision actually needs to be made.
How It Actually Works
The Concept of "Hardening"
In the video, Zapier CEO Wade Foster describes a process called "hardening." In a traditional agentic workflow, the AI decides every single step from scratch every time the process runs. This is where the instability comes from.
Hardening turns these steps into deterministic code. A deterministic step is a rule: "If X happens, always do Y." There is no guessing, no creativity, and no probabilistic variance. As Foster explains, "it is hardened to run a very particular way and then I leave the probabilistic parts to the place where I really value the creativity of these agents."
Example 1: The Granola-to-Drive Workflow
Foster demonstrates a workflow that moves meeting transcripts from a tool called Granola to Google Drive. Instead of letting an agent manage the whole process, the Zap is split:
- Deterministic: Every hour at 15 minutes past the hour, the system fetches new transcripts from Granola.
- Agentic: The AI assesses the transcript against a list of exclusion criteria (e.g., legal strategy, medical matters, or board meetings) to decide if it's too sensitive to share.
- Deterministic: If the AI gives the green light, the system saves the file to Google Drive.
Example 2: The Fantasy Football Manager
In a more complex example, Foster uses a fantasy football workflow integrated with the Sleeper API. The system handles the tedious data gathering deterministically: it fetches the user's roster, the league's rosters, and the waiver wire available players via API calls.
The "managerial" part is where the AI takes over. It analyzes the data to suggest trade partners and waiver wire pickups. Once the report is written, the system switches back to deterministic mode to send the email to a specific address with a locked-down subject line, ensuring the format never changes unexpectedly.
Technical Deep Dive
To understand why this matters, we have to look at the difference between state machines and LLM inference. A deterministic workflow is essentially a finite state machine; the transition from State A to State B is guaranteed. In contrast, an LLM operates on probability distributions (sampling tokens based on weights).
When an agent is "fully agentic," it is essentially running a loop of Observation → Thought → Action. If the "Thought" phase is slightly off, the "Action" might be an unnecessary API call. If this happens in a loop, you get the token hemorrhage described by Andrew Warner, who spent $264 in two days because his agent was obsessively checking websites.
Next Gen Zaps implement a hybrid architecture. By wrapping the LLM in a deterministic shell, Zapier reduces the context window size for each call. Instead of the LLM needing to remember how to call the Sleeper API and how to send an email, it is only fed the data it needs to make a specific decision. This significantly lowers token consumption and eliminates the risk of the agent "hallucinating" a new way to send an email.
Furthermore, Foster introduces Agentically Managed Workflows. This is a self-healing mechanism. If a deterministic step fails (e.g., an API endpoint returns a 500 error), the system doesn't just crash. It triggers a fallback agent that analyzes the error and the system state to find a temporary fix for that specific run, combining the reliability of code with the problem-solving skills of AI.
Why This Matters
For developers and business owners, the biggest barrier to AI adoption isn't capability—it's reliability. You cannot run a business on a system that might randomly decide to change your branding or stop sending emails because it "thought" a different approach was better.
By moving toward "hardened" workflows, we move away from the "chatbot" era and into the "agentic software" era. We stop treating AI as a magic box that does everything and start treating it as a specialized component within a larger, reliable engineering system.
This approach also solves the economic problem of AI. When you remove the need for an LLM to handle mundane tasks like scheduling or data fetching, you drastically reduce your API costs. You stop paying for "world-class brain surgeons" to do data entry, reserving your token budget for the high-value reasoning tasks that actually move the needle.