# We Trust AI Agents That Break Every Day - Shy Ruparel, Temporal

## Executive summary

The talk addresses the significant gap between the daily usage of AI agents (80.8%) and their deployment into reliable enterprise production systems (49.1%). The core challenge is that while agents are trusted (85.5% trust output), they frequently fail (41.1% experience daily issues). The solution presented is Durable Execution, a Temporal concept that ensures workflows survive worker crashes, maintaining state, preventing the re-spending of LLM tokens, and eliminating the need for complex, unrelated crash recovery code. This approach is critical for building reliable, long-running financial workflows.

## Key takeaways

- Usage vs. Production Gap: While 80.8% of surveyed engineers use agents daily, only 49.1% are running them in production, indicating a major reliability hurdle (30% gap).
- Trust vs. Reality: Despite 85.5% of users trusting agent output at least somewhat, 41.1% report hitting agent issues daily or worse, proving that failure must be accounted for in production design.
- Reliability Gap: Seeing vs. Recovering: The top blockers for productivity are state tracking (35.7%) and debugging (30.9%), highlighting that state management is the most critical aspect of enterprise AI deployment.
- Durable Execution Solution: Durable execution allows a workflow to survive process crashes (e.g., timeouts, memory errors). State and progress are maintained in an event history, ensuring that AI activities and tool calls are executed only once, saving costs and complexity.

## Technical details

- Durable Execution: A Temporal concept where code runs to completion even if the worker process fails. State survives, progress survives, and every step is recorded in an event history. This allows the system to resume from the last successful event without re-running activities or re-spending LLM tokens.
- Workflow Structure: Workflows are deterministic, while non-deterministic or side-effecting actions (like LLM calls or bank API calls) are run within dedicated 'Activities.' The human approval or rejection is captured as a 'Signal,' which is recorded in the history and processed even when the worker is offline.
- State Management: Temporal manages workflow state using plain variables (fields), eliminating the need for external databases or caches for state persistence, which was identified as a major blocker.
- LLM Integration: The demo used a local LLM (Qwen 36B/9B) within a payout workflow. Durable execution ensures that the agent's work (e.g., drafting a brief) is retained and not regenerated upon failure, addressing cost concerns (8 out of 10 people cite cost limits).

## Practical implications

- For build engineers, the focus must shift from merely building the agent logic to building the *reliability* around the agent logic.
- Durable execution simplifies complex financial or multi-day processes by abstracting away manual crash recovery code (retry logic, state tables, reconciliation jobs).
- The architecture must treat failure as a guaranteed event, ensuring that state and costs (like LLM tokens) are preserved across worker crashes.

## Topics

AI Agents, Workflow Orchestration, Reliability Engineering, Distributed Systems, LLM Integration, State Management, State of Development Survey, Temporal

Source: https://www.youtube.com/watch?v=mJ7q8CuVmbw